nexusmem 0.3.3 → 0.5.0

package/CHANGELOG.md

@@ -1,287 +1,350 @@
-# Changelog
-
-Notable changes per published version. Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/);
-this project uses [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
-
-Tags were added retroactively on 2026-08-11 and point at the exact commits the npm tarballs were
-built from, matched by publish timestamp: `v0.1.0` → `67a4776`, `v0.1.1` → `809e62c`,
-`v0.1.2` → `b22a3b0`.
-
-## [Unreleased]
-
-## [0.3.3] — 2026-08-16
-
-### Added
-
-- **`nexusmem status` now shows failure→fix chain counts** — `N/M failure(s) resolved (X retry, Y
-  discussion)`, plus a hint to run `sync --link-failures` when failures remain unresolved. The
-  chain feature is this project's most distinctive capability, but was previously invisible to
-  anyone who didn't already know to query `node_links` directly. Backed by the new
-  `getChainStats` in `src/correlate/failure-fix.ts`, which dedupes failures resolved by both
-  heuristics rather than double-counting them.
-
-### Fixed
-
-- **Discussion-bridge heuristic (`sync --link-failures`): corpus-relative boilerplate tokens no
-  longer produce false-positive failure→fix links.** Re-dogfooded at larger scale against a second
-  real project's history: a command made entirely of words that saturate a project's own corpus
-  (e.g. this repo's own name/verbs) could AND-match an unrelated turn that just happened to mention
-  the same words, and bm25 score alone could not separate that from a true positive (measured: the
-  false positive scored *stronger* than two real true positives). `filterBoilerplateTokens` in
-  `src/correlate/failure-fix.ts` drops any token that appears in over 20% of a project's own
-  `conversation_turn`/`session_summary` history before building the match query — measured against
-  real data, not guessed — and skips the discussion-match attempt entirely (rather than falling back
-  unfiltered) when every token turns out to be boilerplate, since a missed link is preferred over a
-  false one for this heuristic. Below 10 discussable nodes the filter is skipped, since frequency
-  isn't a meaningful signal yet on a young project.
-
-## [0.3.2] — 2026-08-16
-
-### Added
-
-- **`mcpName` field in `package.json`**, required by the official MCP registry
-  (registry.modelcontextprotocol.io) to verify that whoever publishes `server.json` under
-  `io.github.yaminbkk/nexusmem` also controls the `nexusmem` npm package itself — the registry
-  rejects a publish attempt otherwise. No behavior change for CLI/MCP users; this is purely a
-  registry-ownership proof.
-- **`list_recent_memory` MCP tool** — chronological listing of a repository's most recently
-  remembered nodes (git commits, diffs, shell commands, docs, conversation, session summaries),
-  newest first. Distinct from `search_memory`: no query, just "what has this project's memory
-  recorded lately" — built for the VS Code extension's sidebar view, which lists rather than
-  searches. Backed by `MemoryStore.listRecentNodes`, reusing the existing `idx_nodes_project_ts`
-  index.
-- **`sync --prune-source <name>` and `sync --prune-stale-shell`** — drop one source's nodes without a
-  full `--rebuild`, which loses history that can't be re-read from disk (the shell tail window, older
-  conversation turns). `--prune-stale-shell` is a shortcut for the three dead pre-hook shell-scrape
-  sources (`shell:pwsh`, `shell:bash`, `shell:zsh`) at once. Dry-run by default — prints the matching
-  count and does nothing until `--yes` is also given, since this is an irreversible full wipe of the
-  named source(s), unlike `--rebuild`'s no-prompt full-project reset. Also sweeps any prior project
-  identity of this same repo (the id a renamed git remote leaves behind after
-  `fix(store): reconcile memory stranded by a changed git remote URL` migrates what it can) — a
-  live-id-only prune could not reach nodes reconciliation deliberately left in place. Exposed on both
-  the CLI and the MCP `sync_project` tool.
-- **Discussion-heuristic failure→fix chains now surface**, tightened to an AND-joined significant-
-  token match instead of the original OR match. Re-verified against this repo's own real database:
-  5/5 discussion links correct (was ~half wrong when it shipped unsurfaced in 0.3.0). Chains now
-  follow across projects in `query --all-projects` too.
-
-### Fixed
-
-- **`sync_project`'s summary no longer contains raw ANSI color codes on Windows.** `runInit`/`runSync`
-  format their output with picocolors for terminal display, and picocolors treats `platform ===
-  'win32'` as sufficient evidence of color support on its own, without checking `isTTY` — correct for
-  a real terminal, wrong for the MCP JSON-RPC channel, which is piped on every platform. Found live: a
-  real MCP client (the VS Code extension's Output channel) rendered the raw escape codes as literal
-  text instead of color. Stripped at the MCP boundary in `syncProject`, leaving the CLI's own terminal
-  output untouched.
-
-## [0.3.1] — 2026-08-15
-
-### Added
-
-- **`scripts/benchmark.ts` (`npm run bench`) — a reproducible end-to-end token-saving benchmark.**
-  Compares `packed.tokensUsed` against two baselines (full-file-read and `git log -p`) for the same
-  files a query's packed nodes touch, over a query set derived mechanically from the corpus itself
-  rather than hand-picked. Used to measure the README's `## What it costs you` numbers against both
-  this repo (62 commits) and a 9,567-commit external corpus (`vitejs/vite`) — see README for the
-  numbers and their methodology caveats.
-
-### Fixed
-
-- **A generic 2-3 letter local model title, "id" as a search token, and one node's chunks flooding a
-  result set — three ranking/retrieval edge cases found by dogfooding, each verified with a red test
-  before the fix.**
-  - Session-summary titles: the local model sometimes wrote a role-framing line ("Role: Lead Systems
-    Engineer...") instead of a summary, in Thai and English alike. The existing generic-title filter
-    didn't catch either language, since it only matched English words like "summary"/"update".
-  - Search: the token `id` alone was prefix-matching unrelated shell commands like
-    `winget install --id ...`, because every query token was OR-ed with no floor and no stopword
-    list, and bm25 gives a rare-but-generic token an inflated score purely from scarcity.
-  - Packing: `conversation_turn` and `doc_section` both chunk one reply or file into several nodes
-    sharing the same timestamp; up to 2 may now appear in one packed result, down from unlimited.
-
-- **A repo's memory could silently split in two if its git remote URL ever changed** (a GitHub
-  account rename, an org transfer). Project identity is derived from the remote URL on purpose — so
-  the same repo re-cloned to a new path or machine keeps sharing memory — but a changed URL on the
-  *same* path minted a new id and stranded every node synced under the old one, invisible to
-  `status`/`query`/MCP from then on. `sync` now detects a prior id already recorded in the repo's own
-  database and reconciles it forward: recomputable node kinds (session summaries, hook-sourced shell
-  history) are migrated under their correct new id and deduplicated against anything already synced;
-  conversation turns, whose identity can't be recomputed, are reassigned in place. Git commits,
-  diffs, and doc sections are left alone — a normal sync already re-derives them completely, so
-  there is nothing to migrate.
-
-## [0.3.0] — 2026-08-13
-
-**Upgrade note.** The first `sync` after upgrading drops every stored embedding and rebuilds it.
-This is not optional and it is not a bug: embeddings now come from Ollama's `/api/embed`, which
-returns L2-normalised vectors, while the previous `/api/embeddings` did not — measured norms of 1.0
-and 20.7 for the same input. `nodes_vec` ranks by Euclidean distance and records no per-row
-provenance, so a corpus holding both would separate by scale rather than by meaning. `sync` says
-what it dropped, nodes are untouched, and BM25 keeps working while the rebuild runs. On this repo
-the rebuild was 833 nodes in one pass, inside a 9-second sync.
-
-### Added
-
-- **Session summaries via a local model** (`sources.session`, opt-in, off by default). Each
-  finished session becomes one distilled `session_summary` node — decisions and their reasons —
-  alongside the raw exchanges. Runs a local Ollama chat model (`qwen2.5:3b` by default); nothing is
-  downloaded automatically and no transcript leaves the machine. New `scan-session` command, with
-  `--dry-run` to print the exact prompt a session would produce without calling the model.
-  Bounded three ways: a session must be quiet for `settleMinutes` (default 30) before it is
-  eligible, the prompt is hashed so an unchanged session never reaches the model again, and
-  `maxSessions` caps how many are summarized per sync. Every exchange is redacted before the model
-  sees it, and the model's own output is redacted again before it is stored.
-- **`sync --embed-limit <n>`** to cap the embedding pass, for when draining the whole backlog is not
-  wanted.
-
-### Changed
-
-- **The embedding pass drains the backlog in one `sync`** instead of stopping after 200 nodes, and
-  sends texts to Ollama in batches of 32 — measured at 4.21x the throughput of one call per node
-  (20.3ms → 4.8ms per node, over 96 real nodes from this repo). Leaving it uncapped is safe because
-  paging walks rowids monotonically, so a node the provider failed on is passed over rather than
-  retried forever, and because three consecutive dead requests end the pass: an Ollama that is not
-  running now costs three requests rather than one timeout per node.
-- **Embeddings carry a provider identity.** Changing the embedding model, or upgrading from a
-  release that recorded no identity, drops the vectors and re-embeds rather than ranking across a
-  mixture.
-- **Ranking priors now share one budget instead of getting one each.** `signal` and `recency` are
-  query-independent, and each was separately capped at overturning a 2× relevance gap. The score
-  multiplies them, so together they could overturn 4× — which is not a corner case but a description
-  of every commit made during an active working day, fresh and high-signal at once. A query about the
-  PowerShell hook returned two unrelated same-day `fix:` commits at ranks 3 and 4 while the section
-  that answered it sat at rank 6. The 2× budget is now the bound on the priors *jointly*, split
-  between them (`signal^0.215 × recency^0.288`, down from `^0.431` and `^0.576`), and a third prior
-  would re-divide the same budget rather than enlarge it. Measured on four real queries against this
-  repository's memory: the answering section rose in three of them — the rationale for "why BM25
-  before vector search" went from rank 4 to rank 1 — and no query's correct top hit was displaced.
-
-### Known limitation
-
-- Session-summary *titles* depend on the model following a fixed output format, and a 3B model often
-  does not. Measured over 14 real sessions, roughly a third came back usable; the rest were
-  conversational preambles, stray bullets, or a bare "Summary of the Session". Those are rejected
-  and the title falls back to the first line of the question that opened the session — always
-  specific, not always elegant. Compliance was worst on long sessions and on transcripts not in
-  English. `sources.session.model` takes a larger model if it matters.
-
-## [0.2.0] — 2026-08-12
-
-**Upgrade note.** Both new sources are on by default, so the first `sync` after upgrading an
-existing project ingests the patches of its 200 most recent commits and starts recording the
-repository in `~/.nexusmem/projects.json`. Set `sources.diff.enabled` to `false` in
-`.nexusmem/config.json` if you would rather not, and `NEXUSMEM_HOME` relocates the user-scoped
-directory. Nothing existing is rewritten or lost.
-
-### Added
-
-- **Diff-level nodes.** Commit patches are now indexed, one node per changed file, so a question
-  about *what the change looked like* reaches the lines themselves rather than the commit message
-  and a `+41/-6` summary. New `code_diff` kind, `diff` source, `scan-diff` preview command, and a
-  `sources.diff` config block. Read by a second `git log --patch` walk with its own cursor: folding
-  it into the existing `--numstat` walk would put patch text and numstat rows in one field, where a
-  diff line reading `-1\t2\tfoo` is indistinguishable from a real file entry.
-  Bounded on purpose — 200 commits on a first sync, 20 files per commit, no merges (their patch
-  exists only in a combined format this parser does not read), and binaries, lockfiles and build
-  output skipped. Patches are redacted with the shape-matching rules only; the key/value rule that
-  serves prose would rewrite `const apiKey = process.env.SERVICE_API_KEY` into a redaction marker.
-- **Cross-project recall.** `query --all-projects` (and `search_memory`'s `allProjects`) searches
-  every repository NexusMem has been run in on this machine, tagging each result with the repository
-  it came from. Databases stay per-repository — a shared global store was rejected for giving up the
-  property that deleting one repo's `.nexusmem/` removes that repo's memory and nothing else — so a
-  plain index at `~/.nexusmem/projects.json`, written by `init` and refreshed by `sync`, is what
-  makes the others findable. New `projects` command lists it; `--prune` forgets entries whose
-  database is gone. A stale or corrupt registry degrades the query, never fails it.
-  Ranking fuses each project's list by rank (RRF) instead of comparing raw BM25 costs, which are
-  computed against their own corpus and are not comparable across databases. The bias this leaves —
-  every project's rank-1 hit is worth the same, so recall favours breadth — is documented rather
-  than hidden.
-- **Query-aware diff excerpts.** A packed summary is ~320 characters and a patch is thousands, so
-  the packer now picks the hunk whose tokens match the query and starts the excerpt at the changed
-  line. Found by dogfooding: "what flags are passed to every git invocation" retrieved the right
-  file and then spent the whole summary on a class definition seventy lines above the answer.
-  Matching splits identifiers on case and underscore boundaries, because `\bretry\b` does not match
-  `RETRY_DELAYS_MS` and a natural-language question otherwise never meets the code it is about.
-- `CHANGELOG.md` now ships inside the npm tarball. npm's always-included list covers `package.json`,
-  `README` and `LICENSE` but not the changelog, so it previously reached GitHub readers only.
-
-### Internal
-
-- The test suite no longer writes to the developer's real `~/.nexusmem`. `sync` records the
-  repository it ingested in the project registry, and the suite syncs temporary repositories in
-  several places, so a green run left seven dead entries behind — found by running `nexusmem
-  projects` after the fact, not by any test. `tests/setup.ts` now points `NEXUSMEM_HOME` at a
-  throwaway directory for the whole suite, and one test fails if that guard is ever removed.
-- `npm run smoke` drives the *packaged* artifact: build, pack, install into a throwaway directory,
-  then run the installed CLI, an end-to-end ingest/query against a fixture repository, and an
-  `initialize` handshake over real stdio. It also audits the manifest `npm publish` would send,
-  which is a different artifact from the tarball. Both defects that ever reached npm users passed a
-  green unit suite first; each is now pinned by a check verified to fail when the defect is
-  reintroduced. Runs in CI on Linux and Windows as its own job.
-
-## [0.1.2] — 2026-08-10
-
-### Fixed
-
-- `nexusmem --version` printed `0.1.0` on 0.1.1. The version string in `src/cli/index.ts` was a
-  literal separate from `package.json`, and the 0.1.1 bump only touched the latter. `src/mcp/server.ts`
-  had the same problem in its `McpServer` constructor, so an MCP client's `initialize` handshake
-  would have reported the same stale version. Both now read the real version through
-  `readOwnVersion()` in `src/core/version.ts`, which resolves `package.json` via `import.meta.url`.
-  Found by running the published package end to end rather than trusting `npm publish --dry-run`
-  and the registry API, neither of which executes a `--version` flag.
-
-The ingestion and retrieval pipeline was never affected — only the two places that report a version
-independently of running a command.
-
-## [0.1.1] — 2026-08-10
-
-### Changed
-
-- README rewritten for someone deciding whether to read the source: what it does, how retrieval
-  scores, what it costs, and where it breaks. `README.md` ships inside the package, so this is a
-  real change to what npm delivers — but no code changed between 0.1.0 and 0.1.1.
-- Documented the ranking flaw the tool found in itself, and the fact that the conversation source
-  in the sample `status` output is opt-in rather than default.
-- Dropped `&&` from the quickstart, which Windows PowerShell 5.1 cannot parse.
-
-## [0.1.0] — 2026-08-10
-
-First public release.
-
-### Added
-
-- **Collectors.** Git history (commit metadata and diff stats, not diff bodies), shell commands with
-  exit codes via an opt-in PowerShell hook, tracked markdown docs via `git ls-files -- '*.md'`, and
-  opt-in assistant transcripts.
-- **Hybrid retrieval.** SQLite FTS5 BM25 and `sqlite-vec` KNN over 768-dim embeddings, fused with
-  reciprocal rank fusion, then ranked by relevance against signal and recency priors and packed into
-  an explicit token budget.
-- **MCP server** over stdio (`nexusmem mcp`) exposing `search_memory`, `sync_project` and
-  `get_status`, for Claude Desktop, Cursor, Windsurf and other MCP clients.
-- **CLI**: `init`, `sync`, `query`, `status`, `mcp`, and `hook install|remove|status`, plus four
-  dry-run previews — `scan-git`, `scan-shell`, `scan-docs`, `scan-conversation` — that write nothing
-  and print the nodes ingestion would create with their signal scores.
-- Content-addressed node ids (`sha256(projectId + kind + naturalKey)`), so `sync` is idempotent and
-  two clones of one repository share a memory namespace.
-- Everything stays on the machine: one SQLite database in WAL mode under `<repo>/.nexusmem/`.
-
-### Notes
-
-- Requires Node **>= 22**. `better-sqlite3` 12.11.1 publishes no prebuilt binary for Node 20 — its
-  prebuilds start at ABI 127 — so a lower floor would have been a promise the package could not keep.
-  Do not lower it without checking upstream prebuilds first.
-- Not done at this release: diff bodies are not indexed, queries are scoped to a single project,
-  there is no local-model summarization pass, and the conversation collector has never been audited
-  for the stale-node bug that was found and fixed in the docs collector.
-
-[Unreleased]: https://github.com/yaminbkk/NexusMem/compare/v0.3.3...HEAD
-[0.3.3]: https://github.com/yaminbkk/NexusMem/compare/v0.3.2...v0.3.3
-[0.3.2]: https://github.com/yaminbkk/NexusMem/compare/v0.3.1...v0.3.2
-[0.3.1]: https://github.com/yaminbkk/NexusMem/compare/v0.3.0...v0.3.1
-[0.3.0]: https://github.com/yaminbkk/NexusMem/compare/v0.2.0...v0.3.0
-[0.2.0]: https://github.com/yaminbkk/NexusMem/compare/v0.1.2...v0.2.0
-[0.1.2]: https://github.com/yaminbkk/NexusMem/compare/v0.1.1...v0.1.2
-[0.1.1]: https://github.com/yaminbkk/NexusMem/compare/v0.1.0...v0.1.1
-[0.1.0]: https://github.com/yaminbkk/NexusMem/releases/tag/v0.1.0
+# Changelog
+
+Notable changes per published version. Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/);
+this project uses [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+Tags were added retroactively on 2026-08-11 and point at the exact commits the npm tarballs were
+built from, matched by publish timestamp: `v0.1.0` → `67a4776`, `v0.1.1` → `809e62c`,
+`v0.1.2` → `b22a3b0`.
+
+## [Unreleased]
+
+## [0.5.0] — 2026-08-17
+
+### Added
+
+- **`nexusmem forget <value>`: permanent, value-keyed deletion.** The finer-grained complement to
+  `sync --prune-source` (which only deletes at the whole-collector-source granularity): forgets one
+  exact string or `--regex` pattern, deleting every node it currently matches *and* writing a
+  standing deny-list entry so the value can never be re-ingested — closing the one gap an external
+  source-level review flagged as its most serious finding: the append-only shell-hook log (and a
+  full transcript re-read) are untouched by any prune, so `sync --rebuild` used to resurrect exactly
+  what had just been deleted. Every removal leaves a hash-only tombstone (never the forgotten
+  content itself — `body`/`title` are stored as sha256 only) and the whole operation writes one
+  `mutation_audit` row, whether or not anything matched. Dry-run by default, matching
+  `--prune-source`'s convention exactly; `--yes` confirms; `--list` shows active entries. Permanent
+  in v0 — no `--remove`, matching the existing irreversible framing of `--prune-source`/`--rebuild`.
+  See `docs/forget-mechanism.md`.
+
+## [0.4.0] — 2026-08-16
+
+### Added
+
+- **`nexusmem status` now warns when prior project identities still hold nodes.** A renamed git
+  remote can leave stale data under the old identity; status reports both the identity and node
+  counts and points to `sync --prune-source <name>` instead of leaving that data discoverable only
+  through a raw SQLite query.
+- **`docs/competitor-comparison.md`: an honest, source-level comparison against `riponcm/projectmem`** (705
+  GitHub stars vs. this project's 8, at time of writing) — a real, working competitor read in full (not
+  judged from its README), covering where the two tools' failure-tracking, import-graph coverage, and
+  token-savings claims genuinely differ. Cites a freshly re-run benchmark (`npm run bench`, not a stale
+  table) alongside projectmem's own `pjm score` constants, read directly from its source. Linked from the
+  README's `## What it costs you` section and a short website section near the existing stats.
+- **`nexusmem hook git install|remove|status`: a real git pre-commit hook.** Installs a marked block into
+  `.git/hooks/pre-commit` that runs `nexusmem precheck` (no `--strict`, so it can never block a commit on
+  its own) before each commit — the automatic counterpart to the advisory `precheck` command. Refuses to
+  touch a pre-existing foreign hook (husky, lint-staged, lefthook, ...) unless `--force` is passed, in which
+  case it appends after the existing content rather than before, so the foreign hook still runs first and
+  keeps deciding whatever it already decided. Idempotent (`nexusmem hook git install` twice is a no-op) and
+  removable cleanly, including restoring a foreign hook to its original content if one was appended onto.
+  Live-verified against a real scratch git repo on Windows: fresh install, a real `git commit` that
+  triggered the hook and printed a correct precheck report, clean removal, and both the refuse-without-force
+  and append-with-force foreign-hook paths.
+- **`nexusmem precheck`: proactive pre-commit warnings.** Checks staged (or `--working`, or explicit `--files`)
+  files against project memory and warns about unresolved past failures and high recent churn *before* you
+  commit — advisory by default (always exits 0; `--strict` turns an unresolved failure into a non-zero exit).
+  Matches a file's own basename tokens against still-unlinked `shell_command` failures (no
+  `resolved_by:*` link from `sync --link-failures`), reusing `filterBoilerplateTokens` from the
+  discussion-bridge heuristic — now exported and parameterized by node kind so it can be corpus-relative
+  against `shell_command` history instead of only `conversation_turn`/`session_summary`. Churn is scoped to
+  `git_commit`-kind `node_files` touches only, so it doesn't double-count the identical touches `code_diff`
+  nodes also record. Deliberately does not yet install as a real git hook (see the module comment in
+  `src/correlate/precheck.ts` for why capture-time `git status` diffing and past-commit correlation were both
+  rejected) — that's a follow-up once this signal has been dogfooded.
+- **JS/TS import-graph edges.** `nexusmem scan-structure` previews (and `sync` now ingests) file→file
+  import relationships across a project's tracked `.ts`/`.tsx`/`.js`/`.jsx` files — a dependency-free
+  regex extractor (`src/structure/extract.ts`) resolves relative `import`/`export ... from`/`require`/
+  dynamic-`import` specifiers against the tracked-path set, correctly rewriting the common TS/ESM
+  `./foo.js` specifier back to its real `foo.ts` source. Stored in a new `file_edges` table (schema
+  v4), replaced wholesale on every sync since edges describe current tree state, not history. Surfaced
+  as a `structure` line in `nexusmem status`; not yet wired into `query` ranking or exposed as an MCP
+  tool — that's a follow-on design question, not this pass's job.
+
+## [0.3.3] — 2026-08-16
+
+### Added
+
+- **`nexusmem status` now shows failure→fix chain counts** — `N/M failure(s) resolved (X retry, Y
+  discussion)`, plus a hint to run `sync --link-failures` when failures remain unresolved. The
+  chain feature is this project's most distinctive capability, but was previously invisible to
+  anyone who didn't already know to query `node_links` directly. Backed by the new
+  `getChainStats` in `src/correlate/failure-fix.ts`, which dedupes failures resolved by both
+  heuristics rather than double-counting them.
+
+### Fixed
+
+- **Discussion-bridge heuristic (`sync --link-failures`): corpus-relative boilerplate tokens no
+  longer produce false-positive failure→fix links.** Re-dogfooded at larger scale against a second
+  real project's history: a command made entirely of words that saturate a project's own corpus
+  (e.g. this repo's own name/verbs) could AND-match an unrelated turn that just happened to mention
+  the same words, and bm25 score alone could not separate that from a true positive (measured: the
+  false positive scored *stronger* than two real true positives). `filterBoilerplateTokens` in
+  `src/correlate/failure-fix.ts` drops any token that appears in over 20% of a project's own
+  `conversation_turn`/`session_summary` history before building the match query — measured against
+  real data, not guessed — and skips the discussion-match attempt entirely (rather than falling back
+  unfiltered) when every token turns out to be boilerplate, since a missed link is preferred over a
+  false one for this heuristic. Below 10 discussable nodes the filter is skipped, since frequency
+  isn't a meaningful signal yet on a young project.
+
+## [0.3.2] — 2026-08-16
+
+### Added
+
+- **`mcpName` field in `package.json`**, required by the official MCP registry
+  (registry.modelcontextprotocol.io) to verify that whoever publishes `server.json` under
+  `io.github.yaminbkk/nexusmem` also controls the `nexusmem` npm package itself — the registry
+  rejects a publish attempt otherwise. No behavior change for CLI/MCP users; this is purely a
+  registry-ownership proof.
+- **`list_recent_memory` MCP tool** — chronological listing of a repository's most recently
+  remembered nodes (git commits, diffs, shell commands, docs, conversation, session summaries),
+  newest first. Distinct from `search_memory`: no query, just "what has this project's memory
+  recorded lately" — built for the VS Code extension's sidebar view, which lists rather than
+  searches. Backed by `MemoryStore.listRecentNodes`, reusing the existing `idx_nodes_project_ts`
+  index.
+- **`sync --prune-source <name>` and `sync --prune-stale-shell`** — drop one source's nodes without a
+  full `--rebuild`, which loses history that can't be re-read from disk (the shell tail window, older
+  conversation turns). `--prune-stale-shell` is a shortcut for the three dead pre-hook shell-scrape
+  sources (`shell:pwsh`, `shell:bash`, `shell:zsh`) at once. Dry-run by default — prints the matching
+  count and does nothing until `--yes` is also given, since this is an irreversible full wipe of the
+  named source(s), unlike `--rebuild`'s no-prompt full-project reset. Also sweeps any prior project
+  identity of this same repo (the id a renamed git remote leaves behind after
+  `fix(store): reconcile memory stranded by a changed git remote URL` migrates what it can) — a
+  live-id-only prune could not reach nodes reconciliation deliberately left in place. Exposed on both
+  the CLI and the MCP `sync_project` tool.
+- **Discussion-heuristic failure→fix chains now surface**, tightened to an AND-joined significant-
+  token match instead of the original OR match. Re-verified against this repo's own real database:
+  5/5 discussion links correct (was ~half wrong when it shipped unsurfaced in 0.3.0). Chains now
+  follow across projects in `query --all-projects` too.
+
+### Fixed
+
+- **`sync_project`'s summary no longer contains raw ANSI color codes on Windows.** `runInit`/`runSync`
+  format their output with picocolors for terminal display, and picocolors treats `platform ===
+  'win32'` as sufficient evidence of color support on its own, without checking `isTTY` — correct for
+  a real terminal, wrong for the MCP JSON-RPC channel, which is piped on every platform. Found live: a
+  real MCP client (the VS Code extension's Output channel) rendered the raw escape codes as literal
+  text instead of color. Stripped at the MCP boundary in `syncProject`, leaving the CLI's own terminal
+  output untouched.
+
+## [0.3.1] — 2026-08-15
+
+### Added
+
+- **`scripts/benchmark.ts` (`npm run bench`) — a reproducible end-to-end token-saving benchmark.**
+  Compares `packed.tokensUsed` against two baselines (full-file-read and `git log -p`) for the same
+  files a query's packed nodes touch, over a query set derived mechanically from the corpus itself
+  rather than hand-picked. Used to measure the README's `## What it costs you` numbers against both
+  this repo (62 commits) and a 9,567-commit external corpus (`vitejs/vite`) — see README for the
+  numbers and their methodology caveats.
+
+### Fixed
+
+- **A generic 2-3 letter local model title, "id" as a search token, and one node's chunks flooding a
+  result set — three ranking/retrieval edge cases found by dogfooding, each verified with a red test
+  before the fix.**
+  - Session-summary titles: the local model sometimes wrote a role-framing line ("Role: Lead Systems
+    Engineer...") instead of a summary, in Thai and English alike. The existing generic-title filter
+    didn't catch either language, since it only matched English words like "summary"/"update".
+  - Search: the token `id` alone was prefix-matching unrelated shell commands like
+    `winget install --id ...`, because every query token was OR-ed with no floor and no stopword
+    list, and bm25 gives a rare-but-generic token an inflated score purely from scarcity.
+  - Packing: `conversation_turn` and `doc_section` both chunk one reply or file into several nodes
+    sharing the same timestamp; up to 2 may now appear in one packed result, down from unlimited.
+
+- **A repo's memory could silently split in two if its git remote URL ever changed** (a GitHub
+  account rename, an org transfer). Project identity is derived from the remote URL on purpose — so
+  the same repo re-cloned to a new path or machine keeps sharing memory — but a changed URL on the
+  *same* path minted a new id and stranded every node synced under the old one, invisible to
+  `status`/`query`/MCP from then on. `sync` now detects a prior id already recorded in the repo's own
+  database and reconciles it forward: recomputable node kinds (session summaries, hook-sourced shell
+  history) are migrated under their correct new id and deduplicated against anything already synced;
+  conversation turns, whose identity can't be recomputed, are reassigned in place. Git commits,
+  diffs, and doc sections are left alone — a normal sync already re-derives them completely, so
+  there is nothing to migrate.
+
+## [0.3.0] — 2026-08-13
+
+**Upgrade note.** The first `sync` after upgrading drops every stored embedding and rebuilds it.
+This is not optional and it is not a bug: embeddings now come from Ollama's `/api/embed`, which
+returns L2-normalised vectors, while the previous `/api/embeddings` did not — measured norms of 1.0
+and 20.7 for the same input. `nodes_vec` ranks by Euclidean distance and records no per-row
+provenance, so a corpus holding both would separate by scale rather than by meaning. `sync` says
+what it dropped, nodes are untouched, and BM25 keeps working while the rebuild runs. On this repo
+the rebuild was 833 nodes in one pass, inside a 9-second sync.
+
+### Added
+
+- **Session summaries via a local model** (`sources.session`, opt-in, off by default). Each
+  finished session becomes one distilled `session_summary` node — decisions and their reasons —
+  alongside the raw exchanges. Runs a local Ollama chat model (`qwen2.5:3b` by default); nothing is
+  downloaded automatically and no transcript leaves the machine. New `scan-session` command, with
+  `--dry-run` to print the exact prompt a session would produce without calling the model.
+  Bounded three ways: a session must be quiet for `settleMinutes` (default 30) before it is
+  eligible, the prompt is hashed so an unchanged session never reaches the model again, and
+  `maxSessions` caps how many are summarized per sync. Every exchange is redacted before the model
+  sees it, and the model's own output is redacted again before it is stored.
+- **`sync --embed-limit <n>`** to cap the embedding pass, for when draining the whole backlog is not
+  wanted.
+
+### Changed
+
+- **The embedding pass drains the backlog in one `sync`** instead of stopping after 200 nodes, and
+  sends texts to Ollama in batches of 32 — measured at 4.21x the throughput of one call per node
+  (20.3ms → 4.8ms per node, over 96 real nodes from this repo). Leaving it uncapped is safe because
+  paging walks rowids monotonically, so a node the provider failed on is passed over rather than
+  retried forever, and because three consecutive dead requests end the pass: an Ollama that is not
+  running now costs three requests rather than one timeout per node.
+- **Embeddings carry a provider identity.** Changing the embedding model, or upgrading from a
+  release that recorded no identity, drops the vectors and re-embeds rather than ranking across a
+  mixture.
+- **Ranking priors now share one budget instead of getting one each.** `signal` and `recency` are
+  query-independent, and each was separately capped at overturning a 2× relevance gap. The score
+  multiplies them, so together they could overturn 4× — which is not a corner case but a description
+  of every commit made during an active working day, fresh and high-signal at once. A query about the
+  PowerShell hook returned two unrelated same-day `fix:` commits at ranks 3 and 4 while the section
+  that answered it sat at rank 6. The 2× budget is now the bound on the priors *jointly*, split
+  between them (`signal^0.215 × recency^0.288`, down from `^0.431` and `^0.576`), and a third prior
+  would re-divide the same budget rather than enlarge it. Measured on four real queries against this
+  repository's memory: the answering section rose in three of them — the rationale for "why BM25
+  before vector search" went from rank 4 to rank 1 — and no query's correct top hit was displaced.
+
+### Known limitation
+
+- Session-summary *titles* depend on the model following a fixed output format, and a 3B model often
+  does not. Measured over 14 real sessions, roughly a third came back usable; the rest were
+  conversational preambles, stray bullets, or a bare "Summary of the Session". Those are rejected
+  and the title falls back to the first line of the question that opened the session — always
+  specific, not always elegant. Compliance was worst on long sessions and on transcripts not in
+  English. `sources.session.model` takes a larger model if it matters.
+
+## [0.2.0] — 2026-08-12
+
+**Upgrade note.** Both new sources are on by default, so the first `sync` after upgrading an
+existing project ingests the patches of its 200 most recent commits and starts recording the
+repository in `~/.nexusmem/projects.json`. Set `sources.diff.enabled` to `false` in
+`.nexusmem/config.json` if you would rather not, and `NEXUSMEM_HOME` relocates the user-scoped
+directory. Nothing existing is rewritten or lost.
+
+### Added
+
+- **Diff-level nodes.** Commit patches are now indexed, one node per changed file, so a question
+  about *what the change looked like* reaches the lines themselves rather than the commit message
+  and a `+41/-6` summary. New `code_diff` kind, `diff` source, `scan-diff` preview command, and a
+  `sources.diff` config block. Read by a second `git log --patch` walk with its own cursor: folding
+  it into the existing `--numstat` walk would put patch text and numstat rows in one field, where a
+  diff line reading `-1\t2\tfoo` is indistinguishable from a real file entry.
+  Bounded on purpose — 200 commits on a first sync, 20 files per commit, no merges (their patch
+  exists only in a combined format this parser does not read), and binaries, lockfiles and build
+  output skipped. Patches are redacted with the shape-matching rules only; the key/value rule that
+  serves prose would rewrite `const apiKey = process.env.SERVICE_API_KEY` into a redaction marker.
+- **Cross-project recall.** `query --all-projects` (and `search_memory`'s `allProjects`) searches
+  every repository NexusMem has been run in on this machine, tagging each result with the repository
+  it came from. Databases stay per-repository — a shared global store was rejected for giving up the
+  property that deleting one repo's `.nexusmem/` removes that repo's memory and nothing else — so a
+  plain index at `~/.nexusmem/projects.json`, written by `init` and refreshed by `sync`, is what
+  makes the others findable. New `projects` command lists it; `--prune` forgets entries whose
+  database is gone. A stale or corrupt registry degrades the query, never fails it.
+  Ranking fuses each project's list by rank (RRF) instead of comparing raw BM25 costs, which are
+  computed against their own corpus and are not comparable across databases. The bias this leaves —
+  every project's rank-1 hit is worth the same, so recall favours breadth — is documented rather
+  than hidden.
+- **Query-aware diff excerpts.** A packed summary is ~320 characters and a patch is thousands, so
+  the packer now picks the hunk whose tokens match the query and starts the excerpt at the changed
+  line. Found by dogfooding: "what flags are passed to every git invocation" retrieved the right
+  file and then spent the whole summary on a class definition seventy lines above the answer.
+  Matching splits identifiers on case and underscore boundaries, because `\bretry\b` does not match
+  `RETRY_DELAYS_MS` and a natural-language question otherwise never meets the code it is about.
+- `CHANGELOG.md` now ships inside the npm tarball. npm's always-included list covers `package.json`,
+  `README` and `LICENSE` but not the changelog, so it previously reached GitHub readers only.
+
+### Internal
+
+- The test suite no longer writes to the developer's real `~/.nexusmem`. `sync` records the
+  repository it ingested in the project registry, and the suite syncs temporary repositories in
+  several places, so a green run left seven dead entries behind — found by running `nexusmem
+  projects` after the fact, not by any test. `tests/setup.ts` now points `NEXUSMEM_HOME` at a
+  throwaway directory for the whole suite, and one test fails if that guard is ever removed.
+- `npm run smoke` drives the *packaged* artifact: build, pack, install into a throwaway directory,
+  then run the installed CLI, an end-to-end ingest/query against a fixture repository, and an
+  `initialize` handshake over real stdio. It also audits the manifest `npm publish` would send,
+  which is a different artifact from the tarball. Both defects that ever reached npm users passed a
+  green unit suite first; each is now pinned by a check verified to fail when the defect is
+  reintroduced. Runs in CI on Linux and Windows as its own job.
+
+## [0.1.2] — 2026-08-10
+
+### Fixed
+
+- `nexusmem --version` printed `0.1.0` on 0.1.1. The version string in `src/cli/index.ts` was a
+  literal separate from `package.json`, and the 0.1.1 bump only touched the latter. `src/mcp/server.ts`
+  had the same problem in its `McpServer` constructor, so an MCP client's `initialize` handshake
+  would have reported the same stale version. Both now read the real version through
+  `readOwnVersion()` in `src/core/version.ts`, which resolves `package.json` via `import.meta.url`.
+  Found by running the published package end to end rather than trusting `npm publish --dry-run`
+  and the registry API, neither of which executes a `--version` flag.
+
+The ingestion and retrieval pipeline was never affected — only the two places that report a version
+independently of running a command.
+
+## [0.1.1] — 2026-08-10
+
+### Changed
+
+- README rewritten for someone deciding whether to read the source: what it does, how retrieval
+  scores, what it costs, and where it breaks. `README.md` ships inside the package, so this is a
+  real change to what npm delivers — but no code changed between 0.1.0 and 0.1.1.
+- Documented the ranking flaw the tool found in itself, and the fact that the conversation source
+  in the sample `status` output is opt-in rather than default.
+- Dropped `&&` from the quickstart, which Windows PowerShell 5.1 cannot parse.
+
+## [0.1.0] — 2026-08-10
+
+First public release.
+
+### Added
+
+- **Collectors.** Git history (commit metadata and diff stats, not diff bodies), shell commands with
+  exit codes via an opt-in PowerShell hook, tracked markdown docs via `git ls-files -- '*.md'`, and
+  opt-in assistant transcripts.
+- **Hybrid retrieval.** SQLite FTS5 BM25 and `sqlite-vec` KNN over 768-dim embeddings, fused with
+  reciprocal rank fusion, then ranked by relevance against signal and recency priors and packed into
+  an explicit token budget.
+- **MCP server** over stdio (`nexusmem mcp`) exposing `search_memory`, `sync_project` and
+  `get_status`, for Claude Desktop, Cursor, Windsurf and other MCP clients.
+- **CLI**: `init`, `sync`, `query`, `status`, `mcp`, and `hook install|remove|status`, plus four
+  dry-run previews — `scan-git`, `scan-shell`, `scan-docs`, `scan-conversation` — that write nothing
+  and print the nodes ingestion would create with their signal scores.
+- Content-addressed node ids (`sha256(projectId + kind + naturalKey)`), so `sync` is idempotent and
+  two clones of one repository share a memory namespace.
+- Everything stays on the machine: one SQLite database in WAL mode under `<repo>/.nexusmem/`.
+
+### Notes
+
+- Requires Node **>= 22**. `better-sqlite3` 12.11.1 publishes no prebuilt binary for Node 20 — its
+  prebuilds start at ABI 127 — so a lower floor would have been a promise the package could not keep.
+  Do not lower it without checking upstream prebuilds first.
+- Not done at this release: diff bodies are not indexed, queries are scoped to a single project,
+  there is no local-model summarization pass, and the conversation collector has never been audited
+  for the stale-node bug that was found and fixed in the docs collector.
+
+[Unreleased]: https://github.com/yaminbkk/NexusMem/compare/v0.5.0...HEAD
+[0.5.0]: https://github.com/yaminbkk/NexusMem/compare/v0.4.0...v0.5.0
+[0.4.0]: https://github.com/yaminbkk/NexusMem/compare/v0.3.3...v0.4.0
+[0.3.3]: https://github.com/yaminbkk/NexusMem/compare/v0.3.2...v0.3.3
+[0.3.2]: https://github.com/yaminbkk/NexusMem/compare/v0.3.1...v0.3.2
+[0.3.1]: https://github.com/yaminbkk/NexusMem/compare/v0.3.0...v0.3.1
+[0.3.0]: https://github.com/yaminbkk/NexusMem/compare/v0.2.0...v0.3.0
+[0.2.0]: https://github.com/yaminbkk/NexusMem/compare/v0.1.2...v0.2.0
+[0.1.2]: https://github.com/yaminbkk/NexusMem/compare/v0.1.1...v0.1.2
+[0.1.1]: https://github.com/yaminbkk/NexusMem/compare/v0.1.0...v0.1.1
+[0.1.0]: https://github.com/yaminbkk/NexusMem/releases/tag/v0.1.0