Skip to content

Releases

0.9.2

Choose a tag to compare

@Newmannn91 Newmannn91 released this 16 Nov 10:00

0.9.2 — 2026-11-16

Test-only fix for the macOS CI failure on 0.9.1. The cold-index test now passes a realpath'd library root, so vault-root Obsidian image refs resolve on a cold index. The 0.9.1 sandbox check in isRegularFile realpaths the file before comparing it against the root; on macOS the temp dir sits under symlinked /var/folders, so a raw root made the check report 'broken'. No source changes. Tests green on Linux (215/215) and Windows (213/213, 2 checks are POSIX-only).

Fixed

  • macOS CI: cold-index test passes a realpath'd library root. Same pattern as the CI #18 fix — the test's two LocalSearchIndex constructions now use await fs.realpath(libDir) for the rootRealDir argument. Production always passes a realpath'd root, so no source change was needed.

0.9.1

Choose a tag to compare

@Newmannn91 Newmannn91 released this 16 Nov 09:33

0.9.1 — 2026-11-16

Bug-fix release. A full-package review of the 0.9.0 code found 30 issues — 6 high, 16 medium, 8 low; 22 confirmed, 8 partially confirmed, 0 false positives. All 30 are fixed here, verified against the code with the build passing and tests green on Linux (215/215) and Windows (213/213, 2 checks are POSIX-only), including 6 new regression tests.

Fixed

  • Search index concurrency. doRefresh no longer races buildAll; the lock heartbeat no longer races releaseLock; a failed buildAll no longer leaves the index reporting ready while it is broken.
  • Sandbox and symlinks. atomicWrite no longer resolves a dangling relative symlink to a file outside the library; sandboxPath validates the path before stat and no longer swallows SandboxError; walkTextFiles blocks symlinks under archive/; archive detection is case-insensitive; the BFS walk visits real directories before symlinked aliases.
  • Tags and front matter. Plain-form quoted # tags (e.g. tags: "a #b") are no longer truncated; a leading UTF-8 BOM no longer breaks front-matter parsing; flow-list comma splitting is fixed.
  • Images. resolveImageRef validates before stat; image URLs with a query string resolve correctly; the image scan skips fenced code blocks and preserves document order.
  • Save and read. Host-config backup no longer overwrites a backup made within the same millisecond and is created atomically; saveFile image names are collision-safe; read_file respects the notes budget; saveFile rejects an existing directory as its target; readFile boundary checks fixed.
  • Search and scope. The direct-scan fallback skips cloud placeholders and dot-directories; scope matching is case-insensitive on macOS/Windows; search errors are logged instead of silently swallowed.
  • Network and system. IPv6 ::ffff:0:0:0/96 is blocked from SSRF checks; downloadImage destroy has an error listener; the Windows viewer quotes ,, ;, = in paths; clipboard reads UTF-8 correctly and reports oversize instead of "empty".
  • Other low-priority fixes. Stat errors are logged instead of swallowed; large downloads spool to disk, preview scaling is bounded, and the file-text cache is capped (LRU, 64 items / 8M chars); CLAUDE_CONFIG_DIR ~ is expanded; moveFile preserves symlinks; index and wizard boundary validation.
  • moveFile same-file hardening. Before unlinking the source, moveFile verifies it is still the same file (device, inode, type, size, mtime); if a sync client replaced it mid-move, the new file is kept and the move still counts as done.
  • Test file: CI macOS realpath fixes merged. The 15 realpath spots from the CI fix (8 rootRealDir, 7 resolveIndexCacheDir) are now in the test file alongside the 6 new regression tests (5 from the 0.9.0 review fixes + 1 moveFile same-file test).

0.9.0

Choose a tag to compare

@Newmannn91 Newmannn91 released this 14 Nov 07:37

Maintenance & archiving release. Part 1: tag editing (parts 2–3 — archive/restore and the two-step plan/apply cleanup — follow in the same release).

Added

  • edit_tags tool. Add, remove, or rename one tag on an existing note, in front matter (tags:, inline or block form) or body (#tag), one operation per call. Only the tag lines change; the rest of the file stays byte-identical. Adding a tag to a note with no front matter creates a minimal ---\ntags: [tag]\n--- block at the top; adding a body tag appends #tag on its own line at the end of the file. The search index updates at once, so the note is searchable by its new tags immediately. Adding a tag that is already there, removing one that is not there, and renaming to a name that already exists are no-ops, not errors. Tag names are case-insensitive.
  • archive_file tool. Move a file into the library top-level archive/ folder (created on demand), preserving its relative path (notes/foo.md becomes archive/notes/foo.md). Refuses to overwrite an existing file at the destination, and rejects directories, missing files, archive/ itself, or files already under archive/. The top-level archive/ folder is excluded from search_files (both indexed and direct-scan fallback), while list_directory still shows archive/ at the root. This also applies to a pre-existing top-level archive/ folder the user already had before upgrading — its notes are no longer searchable (whole-library search and an explicit path: 'archive' search both return nothing), while list_directory still shows archive/ and its contents. On Linux and other case-sensitive filesystems, a folder named Archive/ (or any other capitalisation) is likewise treated as the archive folder, because the archive check is case-insensitive; its notes are not searchable either, and archive_file will not move files inside it.
  • restore_file tool. Move a file from archive/ back to its original path by stripping the leading archive/ segment, creating parent directories as needed. Refuses to overwrite an existing file at the destination and rejects paths outside archive/. Updates the search index immediately so the restored file is searchable again.
  • plan_cleanup tool. Generate a read-only cleanup plan for the library (two-step confirmation). Proposes archiving empty files (0 bytes), duplicate files (keeps the earliest-modified copy), and orphan attachment candidates (unreferenced for over 24h); writes the plan as JSON in the per-library cache dir and returns the plan ID and a human-readable preview. Changes nothing. Dotfiles, dot-directories, and files under archive/ are never proposed. A plan holds at most 200 actions; the preview says so and to regenerate after applying. When the orphan scan is withheld (a note could not be read), the preview says how many notes were skipped. No permanent deletion anywhere in 0.9.0 — empty files are archived, not deleted. Paths that fail the sandbox write check (for example a file name containing ':') are never proposed.
  • memdrive-mcp apply <plan-id> command. Execute a cleanup plan from the command line (the MCP server never starts). Before executing, each action is re-verified: the plan ID must match YYYYMMDDHHmmss-<6 hex>, every path is checked against the sandbox (no absolute paths, .., archive/, or dot-segments) and canonicalized, and a fresh diagnostics scan confirms the reason still holds (empty file still empty, duplicate group still intact, orphan still unreferenced). Actions that no longer hold are skipped with the reason. Plans older than 7 days are rejected. The command exits non-zero when every action is skipped. File hashes are computed by streaming, never by loading a whole file into memory. A single action whose path fails the sandbox write check is skipped on its own instead of refusing the whole plan. Each action now carries a kind ('empty', 'duplicate', or 'orphan'); apply re-validates by kind and rejects unknown kinds. The archive/ path check is case-insensitive, matching the archive detection. Duplicate re-validation also requires the keeper to be a different file from the one being archived and still present on disk; the archive/ check in plan generation is case-insensitive too. Duplicate keeper checks treat an unreadable keeper as missing (the re-check never crashes the apply run); keeper paths are screened against the sandbox path rules first.
  • apply_plan tool. Execute a cleanup plan from chat, by its plan ID. Same re-verification and execution as the memdrive-mcp apply <plan-id> command (paths, SHA-256, and modification times are re-checked; actions that no longer hold are skipped; plans older than 7 days are rejected), but callable by the model in conversation — the user just asks to execute the plan in natural language, no terminal needed. The command-line apply is unchanged and still works.

Fixed

  • Hardening & edge cases. edit_tags preserves a leading UTF-8 BOM; search_files reports out-of-range offsets cleanly, applies file_types before the 200-hit direct-scan cap, and guards background index warm-up against unhandled rejections; save_file escapes ] in Markdown image alt text and cleans up saved images only when the document write itself fails; read_file and edit_tags reject files over 256 MB before loading them into memory.
  • Cross-platform filename validation on write. File names containing characters that are invalid on Windows (notably : and trailing dots or spaces) are rejected on write (save_file, archive_file, and restore_file), while files with such names that already exist in the library can still be read.
  • Archive diagnostics counts archived notes' references. scanDiagnostics now also reads Markdown under archive/, resolving each image reference at both the note's current archived position and its original pre-archive position, so images referenced only by an archived note are no longer falsely listed as orphan candidates.
  • Diagnostics withholds the orphan list when the scan is incomplete. If a note cannot be read (oversize, cloud placeholder, or read error), the report now says so — naming the note and the reason — and withholds the whole orphan list rather than risk a false positive that Part 3 cleanup would delete. A placeholder file is only treated as dataless on macOS (via isDatalessFile), so a genuinely existing file that merely reports zero blocks on other platforms is still read.
  • Diagnostics only withholds the orphan list for unreadable Markdown notes. The scan-incomplete marker is now set only when a Markdown note cannot be read (oversize, cloud placeholder, or read error); a large non-Markdown file (.log, .csv, .json, .txt) no longer hides the orphan list, since only Markdown notes are scanned for image references.
  • Scoped diagnostics report explains out-of-scope skipped notes. When a diagnostics run is scoped to a directory, the report now says when a skipped note lies outside the requested scope — such notes still withhold the orphan list, because a note anywhere in the library may reference an image inside the scope.
  • moveFile no longer overwrites an existing target, and cleans up on failure. moveFile creates the target with fs.link (falling back to copyFile with COPYFILE_EXCL when hard links are unavailable or the move crosses devices) and only then removes the source; if removing the source fails, or a fallback copy fails midway, the just-created target is deleted before the error is rethrown. A copyFile failure with EEXIST leaves the pre-existing target untouched, since it was not created by the move.

0.8.4

Choose a tag to compare

@Newmannn91 Newmannn91 released this 11 Nov 14:48

Tags release: the tag work planned for 0.8.3 lands here — guidance, a first-save tip, and a diagnostics tag list — plus a flow-list quote fix and a trimmed read_file description.

Added

  • Tag guidance in the save_file description. The tool now explains the two tag syntaxes (YAML front matter tags: or body #tag) and when to suggest adding tags: only when the user mentions a category/topic while saving, when the note clearly belongs to a tag already in the library, or when several notes on the same topic are saved in a row (asked once). It never asks for quick screenshot/clipboard saves, when the user only said "save it" with no topic signal, or after already asking once in the conversation. An explicit request to save with tags just saves.
  • One-time tag tip on first save. The first successful save in a library appends a tip about tagging notes; later saves do not repeat it. The marker lives in the per-library cache dir next to the index and is best-effort — a marker write failure never fails the save. The save_file description requires the model to relay the tip to the user when it appears (the server can only ask, not guarantee it).
  • Diagnostics lists tags in use. The diagnostics report ends with a Tags in use (N): section — each tag once, alphabetically, with its file count. It is a descriptive reuse list, not a controlled vocabulary. No tags: Tags in use (0): none.
  • README Tags section. Documents the two syntaxes and the search_files tags filter (a file must carry at least one listed tag to match).

Fixed

  • Flow list with an unpaired single quote. tags: [it's, b] now yields it's and b — the strict bracket regex used to fail and return nothing. The fallback only runs when the strict parse fails and the line contains a ]; a multiline flow list still yields nothing.
  • read_file description trimmed. The TOTAL RESPONSE BUDGET paragraph keeps the three layers (host ~1 MB, server 950,000-byte budget, model pixel-to-token conversion) and drops the v0.1.1 history anecdote and the "other hosts untested" note; it now says read_file returns a budget-fitting preview and to use open_image for the original.

v0.8.3

Choose a tag to compare

@Newmannn91 Newmannn91 released this 11 Nov 05:17

[0.8.3] - 2026-11-10

Bug-fix release (Argon review of 0.8.2 + Muse verification + Sonnet cross-review). The search contract, tools and limits are unchanged.

Fixed

  • Index lock race + heartbeat (Argon #1, high). The lock file now carries a random token; releaseLock only deletes the lock when the file still carries our token, so an old process that wakes up after its lock was stolen no longer deletes the new owner's lock. A heartbeat renews the lock timestamp every 10 s while it is held, so a build running longer than the 30 s staleness timeout cannot have its lock stolen. The acquire side is hardened too (Argon 0.8.3 review): an unreadable lock is only treated as stale when the file itself is old (a just-created empty lock is a live lock being written, not a corrupt one); a stale lock is re-checked immediately before unlinking, so a second process cannot delete the new owner's lock in the steal window; the heartbeat writes atomically (temp file + rename) instead of truncating in place. Round 2: isLockHeldByLiveProcess uses the same mtime rule, so an empty lock file counts as held instead of letting a load run over a just-started build.
  • read_file ignores Obsidian embeds (Argon #2, high). read_file now resolves image references the same way diagnostics does: ![[pic.png]] resolves by filename across the whole library, ![[attachments/pic.png]] resolves vault-root first. .markdown and .mdx are treated as Markdown too (image references resolve; save_file appends references for them). Two gaps from the Argon 0.8.3 review are fixed: on a cold start (or right after a save) every Obsidian embed reported "not found" because the filename map came from an empty/stale index — the resolver now builds/refreshes the index first; and a ref that escapes the library (![[../../outside.png]]) is reported broken instead of probing whether a file exists outside the root (the content was already blocked downstream, but the existence signal alone was a side channel). Round 2: backslashes in a ref (folder\pic.png, synced from Windows) are normalized on every platform; and when another process is still building the index, the resolver waits up to ~2 s for it to finish instead of reporting every embed as missing. Round 3: the non-Obsidian branch in read_file normalizes backslashes too (an inline ![alt](assets\pic.png) used to report ENOENT on Linux/macOS while diagnostics called it valid); the cold-start wait runs at most once per 5 s, so a note with many images cannot block one read_file for 20 s.
  • Filename dates that do not exist (Argon #4). extractDate no longer lets an invalid date roll over (20260231 and 2026-02-31 used to become 2026-03-03); the parsed components must match the source, otherwise the file falls back to its modification time. The filename regex also has digit boundaries, so a longer digit run (order-20261108123.md) is not read as a date. The same rollover check now applies to a front-matter date: (Argon 0.8.3 review), and a trailing YAML comment there is stripped before parsing (date: 2026-10-28 # meeting used to parse as Invalid Date and silently fall back). Round 2: a quoted YAML date (date: "2026-02-31") no longer bypasses the check — the quotes are stripped before matching, since V8 parses the quoted string and rolls it over; years 0–99 are set with setFullYear instead of new Date(y, …) (which maps them to 1900–1999). Round 3: any leading/trailing quote chars are stripped, paired or not — V8 parses even unpaired quotes leniently and rolls over, so leaving one in is worse than stripping it (the bare check below still validates the date).
  • Date bounds parsed in local time (Argon #5). A date-only bound (modified_after: "2026-10-28") is now that calendar day in the server's local timezone, not UTC midnight — in UTC+8 the old code missed files modified that morning. The tool descriptions say so. File dates are now stored on the same basis (Argon #6): a filename/front-matter date is a local calendar day, so a date_from/date_to filter no longer misses it by a day in negative-offset timezones. The on-disk index format version is bumped (2 → 3) so existing indexes rebuild with the new basis.
  • Tag extraction boundaries (Argon #10). A YAML trailing # comment is no longer read as a tag (tags: # comment is empty; - alpha # c1 is alpha); fenced code blocks are skipped (#include in code is not a tag); a body tag must contain at least one letter (#123 is not a tag). Inline code was already excluded (the # there is preceded by a backtick, not whitespace). Argon 0.8.3 review fixes: inline code that contains a space (`git commit #amend`) is now stripped before the tag scan; fenced blocks are stripped on CRLF line endings too (the old regex only matched \n, so a Windows note's code block leaked tags); a comment after tags: no longer eats the block list below it. Sonnet cross-review fixes: a # inside a flow list is a tag character, not a comment (tags: [#a, #b] used to lose b — a 0.8.2 regression; only a # after the closing ] starts a comment); a fence indented up to 3 spaces is still a fence (a fence under a list item used to be missed); an unclosed fence swallows the rest of the note, so a #tag after it is not extracted. Carbon final review fixes: the fence stripper is now a line-by-line CommonMark scan instead of two regexes — a 4+ backtick fence no longer loses the real tags after it (the S3 unclosed-fence cleanup turned the old 4+ backtick tag leak into tag loss, a 0.8.3 regression; the old regex only recognised exactly 3 backticks, so the closing fence never matched and everything after the opening fence was stripped); inline code now matches opening/closing backtick runs of the same length ( foo #bar used to leak #bar); a quoted ] inside a flow list no longer truncates it (tags: ["a]b", c] used to return only a); a multiline flow list is not extracted at all instead of producing a dirty [ tag; addTag rejects only real dirty characters ([, ], {, }, ") — an earlier strict allowlist dropped legitimate front-matter tags with spaces or dots (machine learning, node.js, c++), a 0.8.3 regression caught by Sonnet's final review; it also trims again after stripping quotes/# so a quoted whitespace-only value is not stored as a blank tag (Carbon r2 review).
  • Diagnostics no longer hashes filename-only files (Argon #11, low). .json/.csv/.log files are data, not notes: reading them just to hash for duplicates wasted I/O and reported noise.
  • Linux clipboard: empty is not unsupported (Argon #12, low). When wl-paste or xclip exists but returns nothing, the result is empty, not unsupported (previously both ended in unsupported). Argon 0.8.3 review fixes the headless misreport the other way: a tool that is installed but has no desktop session exits non-zero, so it now reports unsupported instead of empty — the distinction is made from the command's actual exit status, not a which check. Sonnet cross-review: the exit code cannot make that distinction at all — xclip exits 1 both when the clipboard is empty in a real X session ("Error: target STRING not available") and when there is no display ("Error: Can't open display"), so the 0.8.3 fix still reported an empty clipboard as unsupported. The distinction is now made from the stderr text (not available / Nothing is copied = empty; Can't open display / Failed to connect to a Wayland server = no session), verified against a real Xvfb + xclip session. Carbon final review: execFileAsync now sets LC_ALL: 'C' so a non-English locale cannot translate the stderr text the matcher depends on (xclip itself does not localise, but the timeout path leaves stderr empty and the pin costs one line).
  • read_file no longer waits out the full index build on a cold start (Sonnet, medium). resolveNoteImageRef used to await ensureReady() unconditionally, so a first read of a note with an Obsidian embed blocked for the whole build (~2.6 s on a 10k-note library, over a minute at 100k-note scale — past many MCP clients' call timeouts). It now only waits for the build when the index is already ready (a fast refresh); on a cold start it kicks the build off in the background and resolves with what needs no index (vault-root and note-relative paths). Only a basename-only ref (![[pic.png]]) reports broken this once, and a retry after the build finds it. The bounded wait for another process's build (round 2, ~2 s, once per 5 s) is kept, but only when the lock is foreign — waiting for our own background build is the block this removes. Argon round 5: the fire-and-forget change also skipped the fast disk load, so even a perfectly good on-disk index was ignored on the first read after startup (first basename ref broken, later ones fine). Fixed: when no build is in flight and the disk copy is trusted, the resolver first does a fast await load() and uses the index if it lands; only when there is truly no usable index on disk does it fall back to the background build + empty map. Carbon final review: the fast load did not update lastRefreshMs, so the refreshIfNeeded() right after it still synchronously awaited one full scan (292 ms on a 6k-note library). The refresh after a fast load is now a background refresh ({ background: true }); a foreground refresh is only awaited when needsRefresh is set (a file was just saved), matching searchInner.

Changed

  • Incremental refresh test hardening (CI #10 flake, 1/100). The test now gives every write a distinct mtime (seconds apart) so it verifies the refresh logic, not the filesystem's timestamp granularity. An instrumented reproduction script (test/incr-repro.mjs) is kept for diagnosis; the exact trigger of the 1% miss is still undetermined (the same-millisecond mtime hypothesis was ruled out by the earlier reproduction). Same-millisecond...
Read more

0.8.2

Choose a tag to compare

@Newmannn91 Newmannn91 released this 09 Nov 11:27

What's new in 0.8.2

  • Registry metadata only, no behavior change: package.json now carries the mcpName field and the repo ships a server.json, so the server can be published to the official MCP Registry.

Full details in CHANGELOG.md.

0.8.1

Choose a tag to compare

@Newmannn91 Newmannn91 released this 09 Nov 11:09

What's new in 0.8.1

  • First search is faster on large libraries: loading the saved index now takes about half the time it used to.
  • Running two hosts at once is safe now: if Claude Desktop and Cursor are open together, only one builds the index and the other waits, then picks up the finished index. No duplicate builds, no corrupted index.
  • Long paragraphs no longer produce oversized chunks: a long paragraph followed by a big blank gap used to make one chunk bigger than the 600-character cap; chunks now stay within the cap.
  • A damaged index file is caught and rebuilt: if the saved index is broken in a way that would silently miss results, MemDrive now detects it, throws it away, and rebuilds — instead of quietly returning fewer hits.
  • The README now explains keyword search vs exact-phrase search.

Full details in CHANGELOG.md.

0.8.0

Choose a tag to compare

@Newmannn91 Newmannn91 released this 08 Nov 13:43

MemDrive 0.8.0 — Search that actually finds things

This release replaces the old keyword scan with real full-text search. It understands Chinese, ranks results by relevance, and adds filters, pages, and a library health check.

What's new

  • Real search: results are ranked by relevance, and each result shows the passage where your words appear — not just a filename.
  • Chinese that works: searching "说了" finds notes that actually say 说了. Before, it also returned every note containing the single characters 说 or 了.
  • Filters: narrow results by file type, tag, date, or an exact phrase.
  • Results in pages: a search with dozens of hits now comes back a page at a time instead of flooding the chat.
  • New notes are searchable right away: save a note and search for it immediately — no waiting.
  • Library health check: ask for a diagnostics report and see empty notes, duplicates, broken image links, and attachments nothing points to anymore. It only reads — it never changes or deletes anything.
  • More image links found: Obsidian-style embeds (![[img.png]]) and reference-style links are now recognized, alongside the old inline and <img> forms.
  • Works in the background: the search index builds itself on first use. While it's building, search still works the old way. Saving a note updates the index for that note at once.
  • Safe to share one library: using the same library from Claude Desktop and Cursor at the same time no longer risks corrupting the index.

Fixed

  • New notes used to take up to 5 seconds to become searchable. Now they're searchable immediately.
  • A very large log file could freeze the server for ~12 seconds. Now it's skipped (only its name is indexed).
  • Chinese searches no longer return noise from single characters.
  • Image links with # or ? in the path no longer break.
  • Date filters and tag matching had edge-case bugs — fixed.
  • A damaged search index used to make every search fail. Now it rebuilds itself.

Known limitations (coming in 0.8.1)

  • In rare cases a search result passage can run a little long — by the size of the blank space between two paragraphs.
  • Rebuilding the index while another app is using it quietly does nothing; you'll need to retry.
  • If a second app starts while the first is still building the index, it keeps using the old-style search until the next restart.

Verified

  • 125/125 automated tests on Linux, 124/124 on Windows
  • Tested by hand in Claude Desktop, Cursor, and VS Code
  • Speed (10,000 notes): first build ~13–15 s; typical search ~23 ms

0.7.0

Choose a tag to compare

@Newmannn91 Newmannn91 released this 07 Nov 07:25

0.7.0 — Security hardening

  • DNS rebinding protection for image URL downloads (single-resolution, connection-time validation)
  • Atomic writes for notes and host configs (no more half-written files)
  • Stricter image type detection (fixes text files misread as images)
  • Non-ASCII whitespace in filenames no longer breaks image read-back
  • CI now runs typecheck, build, and tests on Windows/macOS/Linux × Node 24/26
  • Full details in CHANGELOG.md

MemDrive 0.6.0

Choose a tag to compare

@Newmannn91 Newmannn91 released this 31 Oct 09:58
9b00773