MinniMemoryMCP changelog

Every published release of the minnimemory package, newest first. Dates are publish dates.

Back to the docs

1.5.0 (2026-11-09)#

The Codex release. 1.4.3 was merged but never published, so everything it held is in this list and npm goes from 1.4.2 to 1.5.0.

  • Fixed (Codex): undo plans the chain from the host the manifest records. With minnimemory registered in both Claude Code and Codex, a plain undo detected Claude Code and restored only the start folder, leaving the user AGENTS.md and the git root's file compiled.
  • Changed (Codex): the server instructions tell the agent to use the four tools, not the shell CLI, and to report doctor's findings instead of editing them away. A live Codex setup session had run the CLI from the shell (38 to 45 calls instead of 4).
  • Fixed (Codex): the CLI undo names every folder it restored. Before, a chain undo printed "restored 1 file(s) in <start folder>" and counted only that folder's deletes, though it had also restored the user AGENTS.md and the git root's file.
  • Changed: one host adapter per host (src/workspace/hosts/); no host-name check outside host.ts and hosts/. No behavior change on Claude Code.
  • Changed (Codex): the size gate. A first compile of an AGENTS.md that Codex loads whole (under its project_doc_max_bytes cap; the user file when it fits inside the cap) is left as it is, and the plan says why: splitting a file that loads whole cost recall in every measured case. Past the cap the compile runs as before; --force still applies it. Claude Code is unchanged.
  • Fixed: a rule stays in AlwaysOnMemory when its section ends in status subsections (### Current state, ### Open items, ### Session log, a changelog). Before, the strict volatility test read the whole section as volatile and routed the rule out with its history, so a rule like "Ask Before Deleting" was no longer loaded; on Codex an agent then dropped a table without asking. The section is now split: the rule stays, and only the status tail goes to __history as one "(continued)" piece, verbatim and in order. A **How to apply:** line under a section, the form Claude Code's memory writes, now marks it as a rule. Every host.
  • Fixed: when the memory hook ranks several targets (the start folder, the chain root, the user file), the one whose best section covers more of the prompt wins, not the first that covers it at all. A __history piece ranks by its own headings instead of last for its file name.
  • Changed: filler words such as "whether", "exactly" and "caused" no longer count as triggers in an OnDemandMemory list. The ranker's stopwords are unchanged.
  • Added (Codex): beside the placed section, the memory hook may place one short companion (at most 200 tokens) when another section covers the prompt as well, plus a closely tied runner-up, a repeat's best lines once per session, and a linked line, each with its own small cap. Claude Code output is unchanged.
  • Changed (Codex): a project AGENTS.md's OnDemandMemory list names the folder its relative paths start from; a prompt of more than 8 distinct terms needs 4 of them in a section to place it, and an always-on match no longer silences the best on-demand section.
  • Fixed (Codex): init's report rule asks for one entry per # AGENTS.md instructions header, user for a header with no folder and files: [] when there is none, and init's tool description on Codex says so. Before, a lone user block was reported as . and an agent with no project file sent nothing.
  • Added (Codex): the memory hook tells the agent once per session to wait on a running test or build with one long poll instead of many short ones. It goes out whatever the memory's size, and with no compile at all. Claude Code is unchanged: its commands block, so it does not poll.
  • Changed (Codex): one apply now covers the whole chain a session loads: the user AGENTS.md, the git root's and the start folder's, each with its own backups and manifest, under one plan and one command; one undo in the start folder puts them all back. The map and the cap total now walk from the git root, so a session in a subfolder counts the root's file. Before, apply compiled only the start folder's file and left the other two as they were. Over MCP, apply and undo list (and with content: true return) every file in the chain. A plan that writes the Codex home says so up front: Codex's sandbox does not let commands write there, so the agent asks to run the command with escalated permissions instead of failing. --budget is the start folder's; every other file keeps its own, so the printed command matches the plan it was printed for.
  • Fixed: init over MCP reads loaded sent as a JSON string, the way agents often send it on their first call. Before, the call was refused and the agent retried, up to four times in live runs. A JSON array of paths is read as files, and on Claude Code a memory folder in it as the memory folder.
  • Fixed (Codex): Codex's memory v2 folder ($CODEX_HOME/memories_v2/, the default for a fresh profile that turns Memories on since Codex 0.165.0) is listed and never compiled, the same as memories/; --memory-dir on it is refused.
  • Fixed: install run from the npx cache (npx -y minnimemory install) no longer registers that copy. npm clears the cache on its own, and the server then showed "Connection closed" while every hook failed with no message. It now writes nothing, exits 2 and prints the durable route: npm i -g minnimemory, then minnimemory install. uninstall still runs from npx. README, ONBOARDING and AGENTS.md lead with those two lines.

1.4.2 (2026-10-29)#

  • Changed: the memory hook places a section only on a strong match. A weaker match on a rare word that names a section prints a one-line pointer to it instead, and anything weaker prints nothing. Chat and generic words (yes, better, option, fix) no longer count toward a match, a short question that refers back to the conversation never gets a whole section, the user CLAUDE.md or AGENTS.md gets a section or nothing, and a section whose text is already in the conversation is not placed again. Offline on two test memories, unneeded placements fell from 8 and 7 of 36 prompts to 1 section and 1 pointer each, with the same workload hits.
  • Fixed: a session started in a subfolder of a compiled project got nothing from the memory hook, because the hook only looked at the folder the session started in. A compiled user CLAUDE.md was never placed at all. init now records, for the folder it runs from, the files the agent reported loaded there, and the hook places from the compiled folders of those files. Started somewhere init never ran, the drift hook asks once per session for init to run there.
  • Fixed: the list in a compiled user CLAUDE.md or AGENTS.md named its routed files by relative path, which inside a project pointed at the project's own file of the same name. It now names them by full path; the user file is machine-local and loads in every project.
  • Changed: when the memory hook places a section, the first one from a file in a session ends with a line naming that file's other sections and their line ranges, so Claude knows the file holds more than the part it was shown.
  • Fixed: on a memory folder in Claude Code's own shape (one file per fact, all indexed), apply keeps the files in place, and the memory hook then saw none of them: it placed nothing and matched the index's own descriptions instead. It now places sections from those files, and the index (which only points at them) never counts as an answer.
  • Faster: the memory hook reads its memory files with one path check per file instead of three, about 90 to 110 ms less per prompt on our test memories. Same results.
  • Added: the memory map init and doctor print now flags an instruction file Claude Code cuts: a subfolder CLAUDE.md or path-scoped rule over 40,000 characters (Claude gets only its start) and any file over 4 MiB (skipped).
  • Changed: the memory hook stays silent when your always-loaded memory already covers the question as well as the section it would place, so the same answer is not added twice. A question about current state or dates still gets its history section.
  • Fixed: when minnimemory is already registered, install now prints the command to change it for the copy you ran (node "<path>" install --force ...), not npx -y minnimemory, which fetches whatever npm has.
  • Fixed: after Claude Code compacts a conversation, the memory hook can place a section again. Before, it remembered placing the section earlier in the session and stayed silent, though compaction had removed it, so Claude had to read the whole memory file instead.
  • Changed: the memory hook stays quiet on short follow-ups about the work in progress ("revert that last edit", "check the logs again", "merge both", "keep developing", "go ahead") and places a one-word prompt only when a section's heading names the word. Before, those turns often placed an unrelated section that then rode along in the conversation for the rest of the session.
  • Added: --profile output, the answer-length rules alone (about 230 always-loaded tokens, a third of full). Output is the most expensive token a session pays for.
  • Changed: the memory hook's placed block is smaller (the header is the section's address only). When the best section is too big to place, the hook places its best part and names the rest by line range instead of only pointing at the file, and never places lines it already placed in the session.
  • Added (opt-in, off by default): install --precise-reads. When Claude reads a whole memory file over about 800 tokens, the Read hook hands back only the sections the current prompt needs and names the others by line range, so Claude can read one of them if it has to. It finds the current prompt in Claude Code's own session transcript; minnimemory stores no prompt text.
  • Added: a usage log and doctor rule MM024. install adds a fourth Claude Code hook, PostToolUse on Read (minnimemory read), that logs which OnDemandMemory files your sessions open (file names and counts only, under your Claude config folder). Once 10 sessions are logged, doctor flags a file Claude read in most of them, since a file nearly every session reads is cheaper always-loaded. Advice only; nothing moves. The hook runs in the background, so Claude never waits for it. Run install again to add the hook.
  • Changed: the memory hook ranks "Current state", open-items and changelog units behind the section that answers, unless the prompt asks about time or state (when, last, changed, now, current, as of, ...). Before, a dense status or changelog line often outranked the section that answers the question, and was placed instead of it.
  • Fixed: a small project host file could lose its largest kept section to OnDemandMemory inside its budget ("kept out to fit the share"), even with <!-- minnimemory:keep --> under the heading, and --budget did not help. A 371-token AGENTS.md of five kept sections lost one under a 2,000-token budget. The file's share was capped at its own size and the compile then charged the title and the OnDemandMemory list against it again. A host file now compiles against the whole budget pool, so sections that fit stay; an explicit --budget below the file still demotes and names a kept section as a rule.
  • Fixed: the memory hook went silent for every memory-folder prompt once the project itself was compiled too: it ranked the compiled project and stopped. It now ranks the project and the memory folder init recorded: the project answers first when it covers the prompt, the memory folder answers the rest (still one unit, once per session, and nothing when the always-loaded text already has it).
  • Changed: a source's task-specific text over about 1,500 tokens is routed as several pieces of about 800 tokens, <parent>__detail-<section>.md, each with its own list line, instead of one <parent>__detail.md, so reading one section no longer loads every module of the file. Smaller sources are unchanged; a workspace compiled before keeps its files until undo and apply.

1.4.1 (2026-10-27)#

    1.4.0 (2026-10-27)#

    • Changed: a first sort of a memory folder that grows what is loaded every turn (the index gets longer, nothing is saved) is now verdict leave: apply refuses it without --force (force: true over MCP) and says how many tokens it would add. A re-apply still sorts.
    • Changed: apply on a memory folder cuts an index hook over the 120 character cap (MM006) to fit, but only when every word of it is in the linked file's frontmatter description:; otherwise the line is kept whole. Undo restores the index byte for byte.
    • Changed: apply on a memory folder swaps the link targets of two index lines that each describe the other's file (MM019); any other MM019 line is still a flag only. MM019 no longer flags a line whose own file holds half its words when the better match has its own line.
    • Added: verdict fix. When a first sort of a memory folder with no imported file is refused because it grows the prefix, apply still fixes the index's own lines (the MM006 cuts and MM019 swaps above) and moves no file, as long as that shrinks the prefix. --force (force: true over MCP) sorts anyway; a folder applied fix-only stays that way on a re-apply, and undo restores the index byte for byte.
    • Not changed: the MCP tool schemas (the host comes from the server's --host, not a per-call argument; measured 2026-10-26: one optional enum on four schemas cost 185 tokens every session). The frozen interface text is untouched.

    1.3.7 (2026-10-22)#

    • Fix (#23): the prompt hook printed nothing when the best memory section was over its 400-token cap. It now prints one pointer line, the file's full path and the section's heading, so the agent learns the file exists when it needs it.
    • Fix (#24): on a memory folder that was never applied, the hook placed sections of files Claude Code already loads (a file your CLAUDE.md imports). The files init recorded as loaded, and their @ imports, now count as always-on, as they already did on a compiled folder.
    • Fix (#29): doctor's MM019 flagged correct index lines on big folders, because a large file held all of their common words. Words now match at the start of a word only, weigh by how few files hold them, and a file over three times the linked file's size counts only its headings. A real 99-file memory folder went from 3 findings to 0; a synthetic test folder with planted wrong descriptions went from 3 to 4, the fourth being a planted one the old rule missed. The finding is low when the linked file still holds half the words.
    • Changed (#30): with a --budget you give to init or doctor, being over it is now the plan's headline ("always-loaded N tokens, M over the B you asked for; the plan below brings it to A") instead of a high MM001 finding, so it no longer counts in the findings or the exit code. A policy's budget and a compiled workspace's own budget still report MM001.
    • Changed (#22): sections that read as hard rules (HARD STOP, never, always or must in the heading or first line) and sections marked <!-- minnimemory:keep --> are placed in the always-loaded text before size decides. One that still does not fit is named in the plan as "a rule". What apply writes changes for files with such sections.
    • Fix (#28): init warns when the folder it looks at is another project's Claude Code memory folder, and drops saved records whose project folder no longer exists.
    • Tests only, no runtime change: the Plugin Sandbox E2E Session Cost Test (tests/integration/e2e/plugin_sandbox_e2e.mjs, live_e2e.mjs --install plugin, selftest fault plugin-no-server).

    1.3.6 (2026-10-19)#

    • Changed: @modelcontextprotocol/sdk 1.30.0 -> 1.32.1 (the directory policy asks local servers to keep dependencies current). No behavior change; the full suite, the MCP client probe and the plugin e2e pass on the new SDK.
    • Plugin: the generated README now defines what "memory" means (your files on disk, never chat history), names the surfaces (Claude Code and Cowork; ignored in claude.ai chat), discloses the bundle's unused install/uninstall commands and the --profile full tools, and gains "Try it on a sample first" and "Security reports" sections; examples/CLAUDE.md ships in the plugin tree; plugin.json carries privacyPolicyUrl, termsOfServiceUrl and author.email. The stale "measured offline, not yet in a live session" line is gone (gate (c) was live).

    1.3.5 (2026-10-17)#

    • Fix: run from the Claude Code plugin (server/cli.mjs), every printed command (apply, undo, the drift notice's remedy, the CLI's own next step) named npx -y minnimemory ..., which fetches npm's latest at run time, from a plugin whose README says it fetches nothing, and could apply a plan hash with a different version than computed it. The tool now recognises its cli.mjs as its own copy and prints node "<plugin root>/server/cli.mjs" ..., the same direct form an installed cli.js already got. Found by the 2026-10-17 plugin audit.
    • Fix: the prompt hook placed a memory unit holding a credential-shaped line (an AWS secret, a Stripe key) into the turn when the prompt asked for it; init flagged the line (MM008) and apply refused it, but the hook had no screen. The hook now runs the MM008 detector over the unit it is about to place and prints one redacted notice instead, once per session.
    • Changed: the prompt hook's per-session record of placed units moved from the shared OS temp folder (<tmp>/minnimemory-hook/, default permissions, predictable name) to <config>/minnimemory/hook-sessions/, beside the load log.
    • Fix: init and doctor were quadratic in the number of memory files: the orphan rule (MM004) rebuilt a regex of every file name for every source file. On one-topic folders of 1,000 and 2,000 files, init took 42 s and 221 s; it now takes about 1 s and 3 s. One index per audit, built once. tests/audit-scale.test.ts holds a ceiling.
    • Plugin: the generated tree carries logo.png (the directory's listing icon) and documentationUrl/supportUrl in its manifest; the plugin's code is readable (not minified) by default, as the directory's checklist asks; scripts/build-plugin.mjs --minified restores the small build. The plugin README lists every file the plugin writes, the exact command apply hands back, and a troubleshooting section.
    • Changed: the LICENSE names the distribution channels (npm, Claude Code plugin marketplaces and their caches, Anthropic's directory) and allows them to store, copy, scan and deliver unmodified copies, which the old clause 2(a) forbade; the text the tool writes into your memory files is licensed to you without restriction as part of your files (the old clause 4 forbade publishing it, which every committed CLAUDE.md did); reading the now-readable plugin source is allowed; the bundled open-source components keep their own licenses (THIRD_PARTY_NOTICES.md); the warranty clause no longer names --allow-write. The grant itself is unchanged: free, unmodified, your own use, commercial use included.
    • Changed: every MCP tool carries readOnlyHint and destructiveHint annotations (Anthropic's directory policy 5.E); init says it writes its discovery record, the other basic tools say they write nothing. The basic profile's tools/list grows from 1,183 to 1,251 tokens (approx-v2).

    1.3.4 (2026-10-14)#

    • Feature: the basic profile registers its four verbs as MCP prompts, so Claude Code lists them as slash commands: /mcp__minnimemory__init, /mcp__minnimemory__doctor, /mcp__minnimemory__apply, /mcp__minnimemory__undo, each with an optional folder argument. A prompt tells the agent to call the matching tool; apply and undo still return the command that runs through your approval. Tools, instructions and footprint are unchanged.

    1.3.3 (2026-10-13)#

    • Fix: the recall hook install registers printed nothing once apply had compiled the memory folder. With no --dir, the hook takes the folder init recorded and always ranked it as an uncompiled folder, which finds no match in a compiled one; it now checks the phase the way --dir does. Effect in a live Claude Code session with the install kept (a 30-task memory benchmark, 2 repeats): +20.3% per session against the unoptimized memory before the fix, -27.3% after.

    1.3.2 (2026-10-13)#

    • Fix (#13): a sorted AlwaysOnMemory/ file that no file Claude Code loads imports any more (the import line was removed, or the user CLAUDE.md is gone) stayed labelled always-on: doctor said "nothing to apply" and the index header still told the agent the file was loaded, so it was never opened. doctor now reports it (MM010, with the import line to put back or the move to OnDemandMemory/ to make), the drift notice says so once with that remedy, and the summary gets a "not loaded:" line. Nothing moves on its own: the import may have gone on purpose (DESIGN.md 6.10, decision 25). Raised only when the loaders are known (a saved set, an import that still reaches the folder, or a host an earlier apply rewrote is in the report), so an incomplete agent report does not raise it.
    • Changed (#14): over MCP, init/doctor on a workspace with nothing to apply lead with that verdict and give the audit as one line per rule (count and worst severity) instead of up to four findings per rule; format: "json" still carries every finding. On an 85-file memory folder the weekly doctor reply went from about 1,400 to about 350 tokens (approx-v2). The terminal output is unchanged.
    • Changed (#14): apply(content: true) no longer sends a backup's content, which is the source file as it is on disk: backups come as copies (from -> to) for the agent to perform before the writes, and files holds the writes only. The JSON form of apply/undo is compact, as doctor's already was. On the same 85-file folder the reply went from about 16,300 to about 14,500 tokens (text) and 26,900 to 24,400 (json).
    • Changed (#14): on a sorted memory folder, doctor's summary now ends with the same "sorted memory folder: N routed files" and "drift: N files changed" lines a project root gets, so day-two drift is not only MM010 lines among the audit findings; and a re-apply whose verdict is "sort" no longer says "skip it if tokens are all you want", since it also brings the index and manifest back in step.

    1.3.1 (2026-10-12)#

    • Fixed: on a memory folder with topic files the index never linked, doctor after apply could keep saying "re-apply" with nothing changed, because choosing the generated list line's trigger words again could move one word. A re-apply now keeps the words an earlier apply wrote for that line, the way it already kept a routed piece's line. A folder sorted by 1.3.0 goes clean on its next doctor with no further apply.
    • Fix (#3): undo in a teammate's clone that had re-applied restored the compiled CLAUDE.md and deleted the routed files it lists, with exit 0. A backup that already holds the OnDemandMemory list is not an original, so undo now refuses and points at git history.
    • Fix (#4): a new memory written into OnDemandMemory/, where the sorted index says new memory files go, made every later apply refuse ("already exists at the destination"). It is now adopted where it is, and a forced undo names it as kept with no index line.
    • Fix (#5): doctor on a memory folder printed apply . as the re-apply command, which run from the project re-applies the project. It now names the folder.
    • Fix (#6): after applying and committing on a branch, main (whose checkout keeps only the gitignored original/) counted as a lost manifest, and the forced undo it offered deleted the branch's backups. It now reads as "not compiled on this checkout": apply says to undo on that branch or merge it first, and undo keeps the backups.
    • Fix (#7): after init on the memory folder, the reply's note about the project read as the next step, and a Haiku session applied the project instead. The note now leads with the folder's own step (apply() with no target), and apply/undo say an omitted target is the recorded memory folder.
    • Fix (#8): init --only-this-folder printed an apply command without the flag, which refused.
    • Fix (#9): init on a memory folder with nothing reported no longer ends with an apply command that would refuse; it says why and how to report.
    • Fix (#10): the drift notice now fires in a session opened in a subfolder of a compiled project.
    • Fix (#11): a deleted .minnimemory/graph.json is written again by the next apply.
    • Fix (#12): after a memory-folder init or apply, the project hint shows only when the project has something to apply. Undo's note for a deleted user CLAUDE.md now says so.

    1.3.0 (2026-10-12)#

    • New: on a memory folder, a topic file the index never linked now gets a generated list line with trigger words, and recall and the prompt hook rank its sections too, so it stops being a dead end the agent can only reach by searching (DESIGN.md 6.10, decision 22). A file whose name marks it as a copy, an old version, a backup or scratch (x (old).md, x.bak.md, x_copy.md, x_v1.md, scratch.md, untitled.md, README.md) is left off the list and named in the plan. The prompt hook also prints only sections that share at least two distinct words with the prompt, and at least half of its words; before, one incidental word was enough to print a block on every such turn.
    • Changed: the prompt hook (minnimemory hook) now places one section, the best, only when it is under 400 tokens, only when the always-loaded text does not hold a better match for the prompt, and only once per Claude Code session (it reads session_id and keeps a small record under the OS temp dir). On a 27-task session it placed about 2.6k tokens instead of about 14.7k (DESIGN.md 6.10, decision 23). recall() gained includeAlwaysOn for the ranking; the recall tool's own output is unchanged.
    • New: doctor --policy <file> checks memory against a team's rule pack (policy.json). Two new rules run only with it: MM020, a line the pack requires is missing from the always-loaded memory (a copy only in an OnDemandMemory file does not count), and MM021, text the pack bans is in a memory file, matched as plain text ignoring case and spacing. Each finding names the pack entry (ref in --json) and never quotes the memory. The pack's always_on_tokens is the MM001 budget when --budget is not given. Patterns are plain text only; a pack with regex is refused and says to list each variant. Without --policy, doctor is unchanged, and the MCP doctor tool does not take a pack.
    • Support window: a workspace compiled by 1.0.0 is now three minor versions back, past the window this release is tested against (6 months or 3 minors). One compiled by 1.2.1 is tested: it reads with no high finding and every routed piece found. Re-apply an older workspace to bring it inside the window.

    1.2.1 (2026-10-09)#

    • New: init and doctor name any loaded CLAUDE.md saved as UTF-16. Claude Code reads such a file as plain text and does not follow its @ imports, so the files it imports never load. Save it as UTF-8.
    • Fixed: a small dated or status section merged with the small sections around it, and the merged file took the "volatile" reason, so rules and how-tos were filed under the __history piece, where an agent looking for a rule does not open. Volatile text now never shares a file with anything else. On the example CLAUDE.md the rules get their own __detail piece: one more list line, so the always-loaded part is 389 tokens, not 363 (Memory Size Test).
    • Fixed: for a memory folder whose path holds $, ` `, ", % or !, the apply and undo commands from Claude Code name the folder ., but the reply did not say to run them from inside it, and run from there they found no init` record. The reply now says where to run, and the record is found by its memory folder.
    • Fixed: a re-apply dropped OnDemandMemory list lines whose file had been deleted, and the whole list when none was left, without a word. It now names each dropped line.
    • Fixed: a manifest.json that is there but was refused (a path it may not hold, a hand edit) was reported as "missing or unreadable". Messages now say which: missing, not valid JSON, or refused.
    • Fixed: npm publish warned that the bin entry was cleaned; it is now written the way npm wants it.
    • New: minnimemory uninstall (the same as install --remove) removes the MCP registration and both hooks, lists the folders still compiled with the undo command for each, and says what it keeps. Documented in the README and the guide.
    • Fixed: install --remove and install treated any node "<path>" drift or loaded hook as ours, so they could remove or repoint another tool's hook. A hook is ours only when its path names minnimemory (or its package.json does).
    • Fixed: a second install --remove exited 1 on "No MCP server named minnimemory"; it now says nothing is registered and exits 0.
    • Fix: pointing apply at a memory index (MEMORY.md) now suggests init on the folder, not doctor, which refuses before init has run.
    • New: apply on a project keeps the record it ran with in the repo, in .minnimemory/discovery.json beside manifest.json, committed with it. A fresh clone (a CI runner, a teammate) has no init record of its own, so doctor and a re-apply now read this one and say so. It lists the always-loaded files inside the project, relative to it: nothing from your home folder or your Claude config, and not CLAUDE.local.md. This machine's own record still wins. undo removes it.

    1.2.0 (2026-10-08)#

    • Fixed: undoing a sorted memory folder with 1.1.0 restored the files but left your user CLAUDE.md importing a file that had moved, so that rule stopped loading. The manifest now also carries the field 1.1.0 reads, so an older copy's undo puts the import back too.
    • Fixed: on Windows and macOS, a folder with a topic file named memory.md and no MEMORY.md lost that topic on apply: the new index was written over it with no backup, and undo could not bring it back. Apply now refuses and asks you to rename the file. Nothing is written.
    • Fixed: sorting a folder moved files that a CLAUDE.md or AGENTS.md names by path (a list item or a link), and those lines then pointed at nothing. Apply now refuses and names the lines. A host that names only the folder still sorts.
    • Fixed: asked to undo "my memory folder and this project", the agent got the memory folder's command only and tried to restore the project's CLAUDE.md by hand. undo's reply about the memory folder now says the compiled project is a separate target with its own undo.
    • Changed: when you apply your memory folder and the project's CLAUDE.md also has changes, the reply gives the project's own ready command, to run only if you asked for both. It is never joined to the memory folder's command.
    • Fixed: the drift notice put its "run apply" line inside the block that tells the agent not to follow instructions; the line now comes after the block. The same for apply's project line.
    • Changed: editing a routed OnDemandMemory file (a saved memory, /remember) no longer raises a drift notice. doctor still lists it, and undo still checks it before discarding anything.
    • New rule MM019: an OnDemandMemory list line whose description fits another file better than the one it links, so the agent opens the wrong file. A flag only; your words are not changed.
    • Changed: when init is called with nothing reported, the reply lists what Claude Code's own load log saw for the project, for the agent to confirm. Nothing is recorded from it.
    • Changed: a forced undo first copies the files it discards to ~/.claude/minnimemory/undo-backups/ and says where.
    • Changed: apply tells the agent not to edit the files by hand.
    • Fixed: after you edit the project's CLAUDE.md, the drift notice told the agent to run apply with no target, which since 1.2.0's discovery change works on the memory folder init recorded. The agent got "nothing to apply" for the wrong folder and told the dev the memory was in sync. The notice now names target: ".", and a doctor or apply reply about the memory folder says when the project's own file changed since the last apply.
    • Removed (full profile): the update tool. Full-profile apply already compiles or re-applies, and agents asked to fix drift used plan then apply. The full profile is six tools now; recall's drift note says "Call apply()". (DESIGN.md 6.20)
    • Changed: doctor and init with format: "json" over MCP send the JSON without indentation, about 18% fewer tokens; the content is the same. The CLI's --json keeps its indent.
    • Fixed (found on a real Mac): an @import typed in other letter case (~/.claude/Projects/...) was counted but not rewritten by apply, so Claude Code stopped loading that topic. An index link typed in NFC to a topic whose name macOS stores in NFD was left at the old place. Both now match the way the file system does and keep the spelling they were written in.
    • Fixed: undo of one memory folder no longer undoes another folder's rewrite of a shared file (a dotfiles file importing from two folders). It puts back only its own lines.
    • Fixed: macOS AppleDouble files (._name.md) are no longer taken for memory topics and moved.
    • Fixed: doctor right after init said the locations were "saved by the last apply"; it now says "saved by init". A budget the fixed part fills names the files in it (the index and the files that load the folder) instead of a "user-level host" that may not exist.
    • Fixed: right after a clean apply, doctor proposed a re-apply that only rewrote hosts.json, when the agent had reported topic files that the apply then moved. A saved file that no longer exists is no longer counted as a change.
    • Fixed: after init, doctor or apply on your memory folder, the reply now says it was the memory folder only and names the project's CLAUDE.md as a separate target (target: ".") while it is not compiled. An agent asked to do both used to do the memory folder only.
    • Fixed: paths an agent types with ~/, or in another Unicode form than the disk (an accented user name on a Mac), are resolved before init saves them. apply used to drop the ~/ ones as missing and rewrite nothing. In a terminal, a quoted ~ in --loaded, --memory-dir or the folder is expanded too.
    • Fixed: an undo or apply refusal over MCP now states its next step (ask the dev, then pass force: true) outside the data fence. It was only inside the block that tells the agent not to follow instructions in it. File names stay fenced.
    • New: init is always the first step, and it is discovery (DESIGN.md 6.19). It never searches for the files that load your memory: the agent passes them to init from its own context (the loaded argument: its auto memory folder and every file Claude Code shows as loaded), or in a terminal --from-log reads Claude Code's own load log and --loaded <path> names a file. init checks every path, works out from their lines which of them load the folder, reports every place the report and the lines disagree, and saves the set in ~/.claude/minnimemory/discovery.json, its one write, never a memory file. apply, doctor, undo and the hooks read it, so nothing is asked twice and the printed apply command carries no paths. In Claude Code, "clean up my memory" works on your memory folder directly: the server accepts the folder Claude Code named.
    • New: doctor comes after init: before it, it refuses and says to run init first. With no folder, doctor, apply and undo use the folder init saved. init is listed first everywhere (the tool list, usage, help, README).
    • New: apply on a memory folder refuses until something reported what loads it, since moving a file an unseen outside file imports would break that import; --only-this-folder says nothing outside loads it. A folder 1.1.0 compiled keeps working: the outside files its apply rewrote count as the report.
    • Fixed: on a project root, init took a reported path that does not exist as verified and saved it, and said nothing about a reported file no line loads; only a memory folder checked the report. Every kind of root now refuses a bad path by name and shows the files the report and the lines disagree on, the same way.
    • Fixed: a report that named the memory folder but no file counted as a report, so apply moved the folder's files and an import of them from the user CLAUDE.md broke (seen with a weak agent that passed an empty list). An empty list is no report now: init says what to pass, apply refuses; --only-this-folder, or the set an earlier apply saved, still says nothing outside loads it.
    • Fixed: an agent that reported only files inside the memory folder left the user CLAUDE.md out, and apply moved the topics it imports. The user CLAUDE.md is now the one file checked without a report: when a line in it loads the folder it is counted, named as not reported, and its import rewritten with the rest.
    • Fixed: apply with no --budget on a compiled workspace planned at the default budget and wrote it into the manifest, so a re-apply after a hand edit changed a 2,300 budget to 2,000 and MM001 appeared. It now keeps the budget the workspace was applied with, like the preview does.
    • Fixed: a project under the home folder counted ~/.claude/CLAUDE.md as one of its own parent-folder files, so doctor . reported the dev's personal memory as the project's. Every command now reads only the folder it is given and the folders below it (DESIGN.md 6.17).
    • Fixed: a file named directly (doctor CLAUDE.md) dropped its own @imports from the total.
    • Fixed: a folder named memory anywhere under ~/.claude/projects/ was taken for a Claude Code memory folder; only projects/<slug>/memory, or the autoMemoryDirectory folder, is one.
    • New: the always-loaded total of a memory folder counts the reported files that load it, through a chain such as CLAUDE.md -> @AGENTS.md, and says what the total covers.
    • A host links a folder by path only where Claude Code would act on it: a list item or a link target. A path in a passing sentence or a code example does not link.
    • New: apply saves the reported set beside the folder's backups, in the gitignored .minnimemory/hosts.json, so a re-apply reads it from anywhere.
    • Fixed: a plain memory folder that a host names by path is read on demand, not counted as always-loaded.
    • Fixed: apply rewrote any @ line whose file name matched a moved file; it now rewrites only lines that resolve into the folder, in every linked host, each with its own backup for undo.
    • Fixed: on Windows and macOS, a path spelled with different letter case (or, on macOS, a different Unicode form) than the file on disk was treated as outside its folder.
    • Fixed: an @import path with accented letters (@C:/Users/José/...) was cut off at the accent.
    • Fixed: a folder opened through a link missed an import that names its real path.
    • Fixed: a Claude Code memory folder's MEMORY.md was counted in full, but Claude Code loads only its first 200 lines / 25 KB. The total now counts that part, the new rule MM017 (high) names the lines past it, and apply refuses a list that would land past it and says how many lines to cut (issue #360).
    • Fixed: over MCP, naming a folder outside the project that nobody reported got a bare refusal; the answer now gives the terminal command for that folder. A ~/ target means the home folder (issues #365, #369).
    • Fixed: MM004 counts a [[name]] link, the form Claude Code's memory uses; files reached that way were reported unreachable (issue #367).
    • New: MM018 (med) reports an OnDemandMemory list line that links a file that does not exist, and names the likely renamed file when one sits beside it (issue #368).
    • Fixed: init on a memory folder opened with a map saying "every session: 0 tokens"; the map now lists what loads with the folder and its total is the audit's (issue #366).
    • Fixed: a "leave this as it is" verdict no longer lists file writes, "-0.0%" and "Run apply()"; it says there is nothing to apply and names the forced write once (issue #370).
    • Fixed: over MCP, "which folder?" now names the server root's path and says a memory folder is outside it, and "nothing to undo" names the folder it checked; an agent undid the project instead of the memory folder and reported it done (issue #371).
    • New: init and doctor take --diff summary (no file diffs), and the terminal command the MCP tools give for a memory folder uses it; the full diff printed 32 KB on a 67-file memory (issue #372).
    • Fixed: install's last line names the same copy it registered, not npx (issue #373).
    • Fixed: an @import chain through a file of any name (~/.claude/CLAUDE.md -> @~/dotfiles/claude/global.md -> a memory topic) was missed, so the topic counted as on demand and apply moved it out from under the import, and Claude Code stopped loading it. Chains are followed through any markdown file, as Claude Code does, and apply rewrites and undo restores the import in that file (issue #374).
    • Fixed: a file that links a memory folder only by loading the next file (a CLAUDE.md that is @AGENTS.md) now says so, loads AGENTS.md, which names this folder (issue #376).
    • Removed: --include-auto-memory, --follow-external-imports and install --auto-memory (still accepted, with a one-line notice). The MCP tools ask for a target instead of assuming the root.

    1.1.0 (2026-10-06)#

    • Known issue, found after publish: the discovery bugs fixed under Unreleased above. A project under the home folder over-counts its always-loaded total by the user-level CLAUDE.md.
    • Kept: the memory-folder list header still says "Read one file, not its siblings." Dropping it made the agent search the routed file instead of opening it, and a measured working session saved 15% instead of 27%.
    • Fixed: an @import written with a Windows drive-letter path (@C:/... or @C:\...) is found like any other import, so the map, the always-loaded total and the audit count the file Claude Code loads. A folder apply rewrites such a line into AlwaysOnMemory/ with the slashes it was written with, and undo restores it.
    • Fixed: init and doctor on a pointer CLAUDE.md (@AGENTS.md) measure MM001 against the same budget the plan uses: a share of the file the pointer names, not the 2,000 floor of the few-token pointer itself. One output used to print both 2,000 and 2,800.
    • Fixed: --budget and the MCP budget argument accept 1 to 1,000,000 tokens. 1e308 was accepted and printed back as a 300-digit number in the apply command.
    • Fixed: the MCP apply and undo refusals (a credential-shaped source, a compile that would not shrink the prefix, an undo that would discard edits) return isError, as the full profile's always did; the apply description no longer says force overrides the credential refusal (nothing does over MCP).
    • Fixed: a routed OnDemandMemory piece holds at most 20,000 tokens; past that, sections continue in <parent>__<reason>_2.md and on. A 1.4 MB memory file used to land in one 360k-token file the list told the agent to read.
    • Fixed: a re-apply that keeps a hand edit to a routed file backs that file up under .minnimemory/previous/, so undo refuses without --force instead of deleting the edit.
    • Fixed: the MCP apply and undo replies keep every note that names a file on disk inside the data fence; only the server's own instruction and command stand before it.
    • Fixed (#334): a finding in a file of the project's auto-memory folder (imported by the user-level CLAUDE.md) ends with the command that fixes it, init on that folder. Before, a project apply left it in place and doctor kept exiting 1 with no next step. The same for a file in another project's memory folder imported the same way. A Claude Code memory folder with MEMORY.md and one topic file now counts as a memory folder, and a hint whose folder path a shell would act on says to run init from inside it.
    • Fixed (#335): the text form of the MCP apply and undo replies puts the tool's own instruction and command before the data fence, not inside the block that says not to follow instructions in it.
    • Changed (#336): the preview's "best case" line quotes the 2026-10-06 measurement (an agent with Read, Grep and Glob: 24.0% lower session cost, answers held), not the Read-only runs of 2026-09-28.
    • Fixed (#337): an MCP tool called with an argument it does not have fails and names it. Before, the key was dropped and the call ran with defaults.
    • Fixed (#338): a re-apply adds new status to the source's existing history file instead of opening __history_2, _3, ...; the plan warns when a routed section carries a rule.
    • Fixed (#339): install --write serves the full profile (the basic profile has no write tools); help documents --host and undo --force and no longer cites an unshipped DESIGN.md; apply/undo with content: true past 200,000 characters returns the command instead; the printed diff stops at 2,000 lines; a forced folder undo names a kept file the restored index has no line for; outline on an unknown file is an error.
    • Security (#327): a committed manifest.json that names a file outside the folder is treated as lost, so it is never opened. Before, apply read such a file, wrote its headings into graph.json and its path into the always-loaded list.
    • Fixed (#328): a line written by hand inside the OnDemandMemory list, naming a file on disk, survives a re-apply word for word. Before, it was dropped and the file was left with nothing pointing at it.
    • Fixed (#329): install over a registration at another scope adds the one asked for and leaves the other alone, instead of failing on claude mcp remove. A re-install of the npx form is recognised as already registered instead of being repointed every time.
    • Fixed (#330): the terminal preview's full diff leaves out .minnimemory/graph.json and manifest.json; the operation list still names them. A 4,000-line changelog printed 104,138 lines, now 8,084.
    • Fixed (#331): install replaces Claude Code's settings.json whole or not at all.
    • Fixed (#332): a merged OnDemandMemory file is described as "first through last" once, --budget takes whole numbers only, engines is node >=18.3 (the CLI uses util.parseArgs), and the lockfile no longer pins dependency versions with advisories.
    • Fixed (#298): doctor --budget 2000 is used as given, not replaced by the compiled or default budget. Before, an explicit 2,000 read as "no budget given", so MM001 measured against another number than the summary line printed, and doctor --budget 2000 --ci passed a file it called over budget. With no --budget, MM001 still reads the compiled budget, and the summary line now prints the budget MM001 used.
    • Fixed (#303): the memory map no longer calls a CLAUDE.md that only mentions AGENTS.md a pointer while listing AGENTS.md as not loaded. The setup block says the host names the file but does not import it, and the not-loaded line says to make CLAUDE.md hold @AGENTS.md.
    • Fixed (#304): apply after a host rename names the renamed backup (CLAUDE.md.bak -> AGENTS.md.bak) and says the host was renamed, over MCP too.
    • Docs (#299): the 40 percent saving holds only above the 2,000-token floor, about 3,300 tokens of host file.
    • Added: the memory map that opens init starts with a setup block: the host file, the file it points at when it is only a pointer, the one file apply rewrites, and whether it is compiled. When the project holds an AGENTS.md that Claude Code does not load, the map ends with a not loaded by Claude Code group naming each one and the reason. init --json and the MCP reply carry both as setup and notLoaded.
    • Fixed: a compiled host file renamed afterwards (CLAUDE.md to AGENTS.md, with or without a pointer left behind) is the same host. Before, doctor reported the old name as missing or edited and said to re-apply, apply answered "nothing to apply", and the two never agreed. Now doctor reports the rename once, one apply carries the manifest and the backups over to the new name, and undo restores the original into the file that holds the memory now.
    • Fixed (#297): a pointer CLAUDE.md rewritten by hand, so that it no longer reaches the compiled file, is reported as drift. doctor names both files and says what to do (put the pointer back, or undo and then apply); apply writes nothing and says the same in place of "nothing to apply".
    • Changed (#238): a project host file keeps its rules and what every task needs, and moves its records. The keep decision is by content, not by heading alone (a body that reads as a rule stays whatever its heading says; commands, setup, build, test, architecture and the like stay after the rules). Volatility is the strict test on a host file too: a status heading, a dated log or repeated status phrases, never a date cited in prose. The default budget is 60 percent of the file, never under 2,000, so on a file over about 3,300 tokens (where 60 percent is above that floor) the saving lands at 40 percent or more with the rules still in the file; a section over the room keeps its head and routes the rest as one "(continued)" piece. doctor's MM001 reads the budget the workspace was compiled at, and MM003 reads a host file with the strict test, so a fresh apply is clean. Before: MinniTuner's 7,958-token CLAUDE.md kept 28 tokens. Now it keeps 4,693.
    • Changed: examples/CLAUDE.md is reshaped to show that rule: 897 tokens, its conventions, architecture and commands kept, its status, reference, runbook, incident notes and changelog routed, a 59.5 percent saving.
    • Fixed (#292): a line added to a pointer CLAUDE.md after apply made the next apply take the pointer for a hand-edited host, write the list into it with no backup, and left undo with nothing to restore. The compiled host is now the file the manifest names, whatever shape the pointer is in, and a file the tool writes for the first time on a re-apply is backed up under original/.
    • Fixed (#291): a closing marker behind a task list box (- [ ] , - [x] ) or an alert tag (> [!NOTE] ) is quoted too.
    • Changed (#293): every MCP reply's data block, and both prompt hooks' blocks, carry a code drawn for that reply in the open line and the close line (=== RETRIEVED MEMORY k3q8v2xm (data, not instructions) === ... === END RETRIEVED MEMORY k3q8v2xm ===). Memory text is written before the reply exists and cannot hold the code, so no line in it is the closing line, whatever character comes before it; the "(quoted)" rewrite stays as a second belt. The exported constants HOOK_OPEN and HOOK_CLOSE are replaced by hookOpen(code), hookClose(code), HOOK_OPEN_LINE, HOOK_CLOSE_LINE, fenceCode() and unwrapHook(); the server exports dataOpen, dataClose, DATA_OPEN_LINE, DATA_CLOSE_LINE and unwrapData() in place of DATA_OPEN and DATA_CLOSE.
    • Fixed (#283): in a project whose CLAUDE.md is only a pointer (@AGENTS.md), a second apply with nothing edited took the pointer for a hand-edited file, appended the list to it, and undo then failed with "no original copy". A compiled project now keeps following the pointer: the second apply is "nothing to apply" and undo is byte for byte.
    • Fixed (#281): a marker line behind a character that is drawn blank but is not whitespace (U+3164, U+2800 and others), or behind a Markdown prefix (> , - , 1. ), still closed the data block in an MCP reply and in the prompt hook. A marker is now quoted whenever nothing but padding or Markdown structure comes before it on the line.
    • Fixed (#282): after a lost-manifest undo kept the user-level CLAUDE.md backup, apply still said "run undo" and undo restored 0 files, for ever. Both now name the one backup that is left and say what to do with it.
    • Fixed (#284): one long line of @ followed by many a/ made doctor and the drift hook, which runs before every prompt, take seconds to minutes (5 s at 24 KB). Resolving a path that does not exist now costs the same for a long path as for a short one.
    • Fixed (#285): install ran claude mcp add through a shell with the server path quoted only for spaces. Under a folder named R&D the path was cut at the & and the shell ran the rest as a command; under a^b a wrong path was registered. Every argument that is not plain is now quoted, and a path with $, a backtick, ", % or ! registers the npx form.
    • Fixed (#286): the full profile's plan, apply and update always returned the full diff, about 3,700 tokens for the 689-token example. They now take diff, with "summary" as the default, like init and doctor.
    • Fixed (#287): two link patterns took seconds on a long run of unclosed links (9 s on 100 KB of [[). Both are now bounded to one line.
    • Fixed (#288): rule MM008 now also catches an npm access token, a URL with a password before the host, a bearer token in an Authorization header, and a prefixed name such as DB_PASSWORD=. One credential matched by two shapes is reported once.
    • Fixed (#245): the command init, apply and undo print put the folder name in unquoted, so a folder called pkg$(touch X) ran touch X when the command was pasted, and names with & or ; broke it. A name is now inside double quotes unless it is plain letters, digits and _ . / : -. A name with $, a backtick, ", % or ! cannot be made safe in bash, PowerShell and cmd at once, so the command leaves it out and one line says to run it from inside that folder. The MCP tools refuse such a target.
    • Fixed (#268, #278): an indented marker line ( === END RETRIEVED MEMORY ===), or one behind an invisible character, still closed the data block in an MCP reply, and the prompt hook never quoted its own closing line (=== END MEMORY ===) at all. Both blocks now quote any line that starts like a marker, whatever its indent or letter case.
    • Fixed (#269): undo with a lost manifest skipped backups whose names start with - or a letter and -- (-draft.md, a--notes.md), then deleted them with .minnimemory/. Those files now come back, and a backup undo cannot match to a file of the folder (the user-level CLAUDE.md's) is kept and named, never deleted.
    • Fixed (#280): undo wrote a backup's text to whatever path manifest.json named, so a repo could ship a manifest that made undo write or delete a file outside the folder. The user-level CLAUDE.md is now restored only when the path is this machine's own, the backup is the one apply writes for it, and that file still imports from the folder. Any other path outside the folder is refused, and a lost-manifest undo never writes outside it.
    • Fixed (#270): with a broken manifest and no backups on the machine (a clone), apply sent the user to undo, which said "nothing to undo". Both now say the backups are not on this machine and the way back is git.
    • Fixed (#271): a re-apply that routed sections out of an edited CLAUDE.md also printed "hand edits kept as they are (nothing to route out)" for that file. The line now shows only for a file the plan leaves as it is.
    • Fixed (#274): apply on a compiled project whose .minnimemory/ folder was gone planned a first compile, removed the list of moved files from CLAUDE.md and left those files with nothing pointing at them. It now refuses, names the files, and changes nothing.
    • Fixed (#275): when apply could not write a file (read-only, locked by another program, full disk) it stopped half way with a stack trace and left backups and moved files behind with no manifest. It now puts back everything that run wrote and prints one clear error.
    • Fixed (#276): a memory file with only blank lines got the verdict "compile" and four bookkeeping files to save 2 tokens, and an empty file printed "break-even -1 tokens". A first compile that moves nothing is now "leave", and the break-even is never under zero.
    • Fixed (#277): --help now lists the loaded command that install adds as a hook.
    • Fixed (#246): with .minnimemory/manifest.json missing or unreadable (a git conflict is enough), apply planned a first compile, backed up the trimmed file over the real original and dropped the list. It now refuses and changes nothing; apply never writes over a backup under original/; doctor names the state once instead of advising a re-apply; and undo (with --force, since it cannot see edits) restores every backup from original/.
    • Fixed (#247): undo listed .claude/ (and .claude/OnDemandMemory/) for deletion even when it held the project's settings or a hand-written file, and told a host that cannot run commands to delete every listed path. A folder is now listed only when everything in it leaves with the undo.
    • Fixed (#248): memory text holding the data block's own closing line ended the block early in every read tool's reply. Such a line is now shown with "(quoted)" and a note; the JSON form is unchanged.
    • Fixed (#249, full profile): update wrote a first compile that apply had refused; it now re-applies only. The server instructions name exactly the tools loaded; plan lost its dead update option; doctor refuses an unknown rule id; apply's description matches what it does.
    • Fixed (#250): one routed file deleted by hand made recall fail for every question. It is now skipped and named, modules marks it missing, and the advice says to re-apply, not "rerun init".
    • Changed (#251): undo, MM011 and the README now say plainly that undo works only on the machine that ran apply (the backups are gitignored); on a clone the way back is git history.
    • Fixed (#252): a re-apply's plan said "saves -2 tokens" and "costs 7,924 more than today". It now says "verdict: re-apply" with the before and after size, and no break-even.
    • Fixed (#253): a re-apply after a hand edit backed the host file up twice.
    • Fixed (#254): a very small file gained a made-up "<title> overview" heading, a "kept out" line for a section that stayed, and a negative break-even. None of the three appear now.
    • Fixed (#255): MCP replies named terminal flags and commands (--budget, --force, --allow-secrets, node ... apply); they now name the tool's options and apply(). The terminal output is unchanged. The auto-memory install hint is built in commands.ts.
    • Added (#256): init with format: "json" (MCP) and init --json (terminal) carry the memory map under map.
    • Fixed (#257): content: true's text form used a three-backtick fence for every file, so a file with its own code block broke the split. Each fence is now longer than any run inside.
    • Changed: the preview no longer uses one label for two totals. doctor's check says all always-loaded files; the plan says always-loaded, files this plan rewrites. Scripts that read the old always-loaded prefix: line need the new text.
    • Changed: a compile plan now says its saving is a best case (kept only on turns that open none of the moved files) and cites the measured sessions.
    • Added: a plan warning when a file keeps under a tenth of its tokens, so rules going out with the rest are visible. Routing itself is unchanged.
    • Fixed: the minnimemory command printed nothing on Linux and macOS when run through npm's symlink (npx -y minnimemory --version). It now compares real paths. install registers the real file, not the symlink. 1.0.0 has this bug, so it needs a new release.
    • Fixed: MM002 no longer calls a table of paths and commands a directory listing.
    • Fixed: apply drops list keywords from the end so the lines it writes stay under the MM006 cap (120 characters, one keyword at least). Lines already under the cap are unchanged.
    • Fixed: diff headers no longer show a//home/... for an absolute path.
    • Added (tests only, no change to the package): the compatibility test. Each release saves its compiled sample repo as a fixture (scripts/save-compat-fixture.mjs, run on the bundled build), and tests/compat.test.ts checks this build still reads every fixture inside the support window of 6 months or 3 minor versions: doctor has no high finding, and every routed piece is still listed and found. A release that breaks one must be a major. The Minni orchestrator's launcher relies on it (Suite/MinniHQ/docs/DESIGN.md 0.9). The fixture for 1.0.0 is saved.

    1.0.0 (2026-09-27)#

    The first stable release. Prepared 2026-09-22, published 2026-09-27 as latest; 1.0.0-beta.2 (2026-09-18) stays on the beta tag.

    • Changed (2026-09-27): init opens with the map of every file Claude Code loads for the project, from Claude Code's documented rules: managed policy, user CLAUDE.md and ~/.claude/rules/, CLAUDE.md / .claude/CLAUDE.md / CLAUDE.local.md here and in every folder above, AGENTS.md when no CLAUDE.md exists, recursive and paths:-scoped rules, subfolder files, imports and auto memory, each with when it loads, whose it is and its tokens. --map prints only the map.
    • Added: install registers an InstructionsLoaded hook (minnimemory loaded) that records what Claude Code actually loads per project, machine-local; init checks its map against it and doctor reports a file Claude loaded that the map missed (MM016). install --remove removes it.
    • Fixed: the always-loaded total now counts user rules, the managed policy file, parent folders' CLAUDE files and nested rules, and no longer counts paths:-scoped rules; a host at .claude/CLAUDE.md is found and compiled with its manifest at the project root.
    • Fixed (2026-09-27, macOS dev run): the auto-memory folder and the load log are found the way Claude Code keys them. A project reached through a symlink (macOS /tmp and /var point into /private, or a ~/dev linked to another drive) is keyed by its real path, with the path as typed as a fallback; a git worktree, and any folder inside it, shares the main checkout's folder, as the Claude Code memory docs say. Before, both missed that memory. doctor's note on a repo's own autoMemoryDirectory now says Claude Code does use it in a trusted folder and that minnimemory still does not follow it (the 2026-09-26 security decision stands).
    • Fixed (2026-09-27, end-to-end dev run): an @path import line in a section apply routes out now stays in the always-loaded file, under ## Imports. Claude Code follows imports only from files it loads at session start, so a routed import had silently stopped its file loading.
    • Fixed (2026-09-27, hands-on audit of the map): AGENTS.md is read when the only CLAUDE file above is the user's own ~/.claude/CLAUDE.md or an excluded one (the map was empty for an AGENTS.md repo under the home folder). Printed commands use forward slashes, so Git Bash no longer strips a Windows target path. The map says when its subfolder scan stopped at its cap; doctor and the drift hook skip that scan, and MM016 builds the map only once a load log exists (doctor on a 6,000-folder tree: 4.3 s to 0.2 s). loaded resolves a relative path against the session's folder, and the map ignores logged files that no longer exist.
    • Fixed (2026-09-27, messy-developer live proof): a hand edit apply has nothing to route out of (a sorted rule file, or MEMORY.md outside its list) is kept byte for byte and taken as the new baseline, so the MM010 drift it raised clears; before, apply said "nothing to apply" and the drift never cleared. The accepted file is backed up under previous/, so undo still names it. undo with force: true over MCP returns a command that carries --force.
    • Fixed (2026-09-27, pre-publish audit): a project host file's budget is now that file's alone. The user-level ~/.claude/CLAUDE.md, its imports, the auto-memory index and .claude/rules stay in MM001 (the audit line) but no longer enter the plan's before/after, share or manifest, so a machine with a large user-level memory no longer compiles every project CLAUDE.md down to its title line, and the MCP server's plan (made without the user-level host unless launched with --include-auto-memory) hashes the same as the CLI's, so the printed apply --plan runs.
    • Fixed: install quotes the arguments it hands the shell on Windows, so a cli path with a space (C:\Users\First Last\...) registers as one argument and the server starts.
    • Fixed: a source that is not valid UTF-8 is refused instead of rewritten (and backed up) with replacement characters; a CRLF source is rewritten with CRLF.
    • Fixed: undo refuses without --force when a compiled file was edited since the last apply or the workspace was re-applied since the first, naming what it would discard.
    • Fixed: MM008 also recognises sk-proj- OpenAI keys, AWS secret access keys and plain password:/api_key= assignments that look like a real value.
    • Fixed: .minnimemory/.gitignore (replacing original/.gitignore) keeps original/, previous/ and the drift stamp out of git; a re-apply of an older workspace adds it.
    • Fixed: the memory-folder plan no longer says "nothing here is volatile" beside a routed history piece, or offers a budget that saves a negative number; the CLI's dry-run line prints the apply command on its own line; install warns when it registers a copy in the npx cache; a --budget -5 error says how to write it.
    • Changed (2026-09-27, workflow audit): over MCP, apply and undo return the command to run (npx -y minnimemory apply --plan <hash> [target], npx -y minnimemory undo [target]) and the operation list with paths and line deltas, never file content, unless called with content: true (for a host that cannot run commands). An apply on the example file went from about 3,500 tokens to about 240; on an 85-file memory folder from about 13,400 to about 300 (moves are grouped by destination). The command names the target whenever it is not the server root (absolute for the auto-memory folder) and always uses the npx form, so it runs as printed; the CLI's init and doctor print the same real command (init used to print a literal <hash>).
    • New (2026-09-27): verdict sort. A memory-folder plan that recompiles no always-loaded file and only moves loose topic files says so, with what the sort costs or saves per turn, instead of "compile ... saves -173 tokens".
    • Changed (2026-09-27, MCP live proof): the operations block prints paths relative to the root the header names, groups backups and moves per destination folder, and doctor/init over MCP show four findings per rule (format: "json" keeps every one). On a real 78-file memory folder a summary doctor went from about 11,000 tokens to about 3,500 and apply from about 3,300 to about 900. The CLI prints the same shorter block.
    • Fixed (2026-09-27, MCP live proof): apply is a fixed point. The command apply returns (and the CLI's next-step line) carries the budget the plan was made with, --force on a leave verdict and the CLI's own --profile/--episodic-json, so the hash is recomputed from the same plan (a real agent got "the plan on disk no longer matches" for a plan made at --budget 12000). The plan's after now counts the user-level CLAUDE.md import lines growing by AlwaysOnMemory/. The manifest records the budget it was applied with, and a re-apply with no --budget or the same one leaves every undrifted compiled source alone instead of re-sharing the pool over the trimmed sizes and trimming again; a different --budget recompiles.
    • Fixed (2026-09-27, MCP live proof): install writes the user-scope drift hook into the Claude Code config dir's settings.json (CLAUDE_CONFIG_DIR when set), not always ~/.claude; under a relocated config dir it wrote the hook into the wrong file. The commands the tool prints call the installed copy directly (node "<path>/cli.js" apply ...) when the CLI knows where it runs from, so the plan hash is recomputed by the build that made it; npx -y minnimemory is the fallback.
    • Fixed (2026-09-27, live proof): a memory saved into a topic file the sort only moved is no longer reported as drift by doctor or the drift hook; MM010 checks the files the tool rewrote and the routed pieces, as DESIGN.md 6.11 decision 3 always said. The sorted index's header now names the folder its paths are relative to (in full for the machine-local Claude Code memory folder), after the live proof's agent resolved OnDemandMemory/pc_setup.md against the project directory first and wasted two reads.
    • Fixed (2026-09-27): the drift hook now also checks the project's Claude Code auto-memory folder, whose .minnimemory/ lives outside the project; it never fired there before. The hook costs about 0.15 s per prompt (was 0.35 s): the CLI loads the MCP SDK only for the mcp command.
    • Changed (2026-09-27): install registers the MCP server as a direct call to the installed copy (node "<path>/cli.js" mcp), the same form as the drift hook, so Claude Code does not re-resolve the package through npx on every launch; a registration that runs a different command is repointed on the next install without --force. npx -y minnimemory mcp stays the form for other hosts' JSON snippets and when install cannot tell where it runs from.
    • Fixed (2026-09-27): apply renders its report before writing, so every write shows its real line delta (was +0 -0) and an empty diff: header no longer appears; it ends with what it wrote and the undo command. doctor's drift lines and the MCP error for an unflagged auto-memory target name the exact command to run next.
    • First stable release. Same four verbs as the betas: doctor, init, apply, undo, in the terminal and over MCP (--profile basic, the default); the seven-tool --profile full is unchanged.
    • New: init and doctor are one preview. Both print the audit, then a unified diff of every file apply would write, then the next step ("run apply", or "nothing to apply"). doctor keeps its CI exit code. init is the first-time name, doctor the returning one.
    • New: memory folders compile. apply/undo on an index plus topic files, or on a project's Claude Code auto-memory folder, used to answer "nothing to compile"; now every always-loaded topic file the index (or the user-level host) imports sorts into AlwaysOnMemory/, trimmed in place; every other topic file sorts into OnDemandMemory/, untouched; routed pieces join it there, one piece per parent per reason (<parent>__history.md, <parent>__detail.md), not one per section. A project root goes through the same engine: the host file is trimmed in place, routed pieces land in .claude/OnDemandMemory/. undo puts every file back byte for byte, including a workspace compiled by the pre-reorganization layout (manifest v4/v5), which doctor now names as such and apply/init refuse ("run undo, then apply") rather than compile over it. DESIGN.md 6.10.
    • New: apply returns operations (copy, move, write, delete) instead of full file content for a backup or a sorted move, and carries a plan hash plus the exact npx -y minnimemory apply --plan <hash> [target] command; the CLI's apply gains --plan <hash> and refuses when the plan on disk no longer matches. Drift (MM010) tracks only what the tool wrote: a file saved into AlwaysOnMemory//OnDemandMemory/ by hand is not drift. DESIGN.md 6.11.
    • New: doctor/init take diff: "summary" (default over MCP) or "full": the plan table plus one line per file (action, path, lines added/removed), with the import rewrite and every move/ delete still shown in full. The CLI keeps the full diff by default. DESIGN.md 6.10.
    • New (revised 2026-09-27): on a memory folder the import is the placement decision. Every stable section of an imported file is a candidate to stay, whatever its heading says; "volatile" means a status heading, a changelog-shaped body or repeated status phrases, never a date cited in a rule's rationale (doctor's MM003 reads imported files the same way). With no --budget every stable section stays and only volatile sections move; when nothing is volatile the plan says so and names what a lower budget would save, from a real second plan. A lower --budget is shared fairly (a file that fits an equal share stays whole, the largest files split the rest) and a heading-less file over its share keeps its leading paragraphs and routes the rest as <name>__detail.md. A single host file keeps the flat 2,000 default. The first cut (a proportional share of half the sources, hint-matched headings, the lenient volatility scorer) hollowed the maintainer's own rule files to frontmatter shells. DESIGN.md 6.10, decision 8.
    • New: the always-loaded count includes .claude/rules/*.md and CLAUDE.local.md, which Claude Code also auto-loads beside the host file, alongside the user-level ~/.claude/CLAUDE.md and its imports (behind --include-auto-memory over MCP; always in the terminal).
    • Changed: when a project has no memory file but its auto-memory folder exists and was not included, the error names the folder and the flag instead of listing places it did not look.
    • Changed: over MCP, apply and undo accept the auto-memory target behind --include-auto-memory, like doctor and init; paths outside the server root come back absolute, with the folder named.
    • Changed: undo gains --force (force: true over MCP) for a memory folder whose routed files were edited since the last apply.
    • Changed (2026-09-27): install writes the drift hook as a direct call to the installed CLI (node "<path>/cli.js" drift) instead of npx -y minnimemory drift: about 0.3 s per prompt instead of 1.3 to 1.7 s. An existing npx entry is updated in place; re-run install after an upgrade. init()'s MCP description is one line (basic tools/list 986 -> 940 tokens). The plan's verdict line states the per-turn arithmetic: the saving on a turn with no routed file open and the extra cost on a turn with every routed file open.
    • Fixed (2026-09-27): doctor, apply and undo given the host file's path on a compiled workspace now read that workspace (they used to see a bare file, plan to strip its fenced list, and find nothing to undo). apply's write guard for a memory folder allows exactly the user-level host file outside the folder, not its whole directory. The MEMORY.md-as-host message points at doctor/apply on the folder. Dead write-protocol module and five unreachable exports removed. README, GUIDE and the site transcript regenerated from a real 1.0.0 run.
    • New: minnimemory drift, a Claude Code UserPromptSubmit hook that prints one short notice when a memory file changed since the last apply, once per change, nothing when clean. minnimemory install turns it on by default now; --no-hook opts out.
    • New: minnimemory install --remove removes the MCP registration and the drift hook install added, leaving every other key and every other hook in Claude Code's settings untouched; every successful install path prints the next step in one line.
    • New: the apply refusal on a source too small to benefit (verdict "leave"), in the CLI and over MCP, shows the always-loaded token count before and after, and the break-even figure.
    • Removed: the compiler no longer adds section anchors to the OnDemandMemory list. They raised the always-loaded prefix on every turn against a saving that was only ever projected, and no freshly compiled workspace could produce one. Compiled output is otherwise unchanged. doctor's MM014 stays, so a workspace compiled by an earlier version still gets told what its anchor lines cost.
    • The package ships as a minified bundle (pkg/) with one public declaration file. Behaviour is unchanged.
    • The generated instruction block carries a one-line MinniAI copyright notice, and the LICENSE gains a clause on the generated instructions. The LICENSE no longer calls itself a beta license.
    • Claude Code is the only supported host; the README no longer documents other hosts.
    • README rewritten terminal-first: install, first doctor, apply, undo, then use, commands, security and license.

    1.0.0-beta.2 (2026-09-18)#

    • minnimemory install registers the server with Claude Code in one command; running the bare package with no TTY starts the server, so MCP hosts can launch it directly.
    • Shipped comments and strings scrubbed of internal paths; a leak check runs before every publish.

    1.0.0-beta.1 (2026-09-17)#

    • First public release on npm.