# Changelog

All notable changes to apsolut-cortex are documented here.

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

## [Unreleased]

## [0.12.11] – 2026-07-09

### Fixed
- **Windows Git Bash diagnosis was giving wrong advice.** When `usr\bin\bash.exe` was absent, `doctor`/`init`/README recommended pointing `CLAUDE_CODE_GIT_BASH_PATH` at `<git>\bin\bash.exe` — but that stub is only a ~47 KB wrapper that re-execs the *missing* `usr\bin\bash.exe`, so it fails identically. The recommendation was broken in exactly the case it was meant to fix. `usr\bin\bash.exe` is now treated as the single source of truth: if it's absent the install is diagnosed as **partial/MinGit** and the remedy is to install the full Git for Windows (`winget install --id Git.Git -e` or git-scm.com), never the wrapper.
- **`doctor` now verifies bash by running it**, not by checking the file exists. It executes the resolved bash (`bash -c "exit 0"`), so a broken wrapper — which exists on disk but won't exec — is correctly reported as broken instead of "✓ set and exists". This catches the "looks configured but hooks still fail" state that a file check can't.
- README troubleshooting rewritten: clarifies Claude Code runs command hooks through Git Bash (not cmd/PowerShell/WSL), that a partial Git is the usual cause, and that pointing at the `bin\` wrapper won't help.

## [0.12.10] – 2026-07-09

### Added
- **`apsolut-cortex doctor`** — diagnoses why hooks aren't firing. First check targets the Windows Git Bash issue below: a broken hook can't report itself, so a CLI-side check is the only place to catch it. Prints the exact `~/.claude/settings.json` line (with the correct, verified path) to paste. Detection lives in `src/gitbash.ts` as a pure function over an injectable file-existence probe, unit-tested against both git layouts (slim and full) since the failure can't be reproduced on a machine with full Git.

### Fixed
- **Keyring native dependency can no longer crash the whole tool.** `@napi-rs/keyring`'s loader requires a platform `.node` binding at import and *throws* if it can't load — a missing `libsecret-1.so.0` on minimal Linux/containers, an arch outside its prebuild matrix, or the npm optional-dependencies bug (npm/cli#4828). Because `db.ts` statically imported the keyring, that throw propagated during module evaluation and crashed **every hook, the MCP server, and the CLI (including `doctor`) at startup** — on the exact platforms where it bites, before any `try/catch` could run. This defeated the intended plaintext fallback. The binding now loads through a guarded `createRequire`, so an unavailable backend degrades to "unencrypted" (the opt-in default) instead of taking the process down. `setDbKey` still fails loud when a user explicitly enables encryption on a backend-less install; a genuinely encrypted DB still fails safe with `SQLITE_NOTADB` rather than being silently overwritten. Encryption remains experimental/opt-in and is treated as a dormant future feature on all platforms for now.
- **Windows: hooks silently no-op on slim/MinGit Git installs.** Claude Code runs command-type hooks through Git Bash, resolving it by default as `<git>\usr\bin\bash.exe`. On a slim / MinGit-style install only `<git>\bin\bash.exe` exists, so Claude Code can't launch bash and every cortex hook (SessionStart, PostToolUse, Stop, SessionEnd, …) fails with a *non-blocking* error — memory capture, session-start, etc. all quietly stop working with no obvious failure. Not a bug in cortex's own code, but cortex is what breaks. `init` now detects the slim layout at install time and prints the remedy (set `CLAUDE_CODE_GIT_BASH_PATH`); `doctor` detects it on demand. Instruct-only by design — neither command mutates the user's global `settings.json` `env` block. README gains a "Windows: hooks not firing?" troubleshooting section.

## [0.12.8] – 2026-07-04

Fixes from the 2026-07-03 pre-release deep review (P1 batch).

### Fixed
- **Compression failure is now loud and non-destructive.** Provider failure (no `ANTHROPIC_API_KEY`, Ollama unreachable, or circuit breaker open) was previously indistinguishable from "nothing worth keeping": SessionEnd marked observations promoted and the in-session worker advanced the compression cursor, permanently consuming captured work with zero memories extracted. `compressSession` now returns `null` on failure; observations stay unprocessed, the cursor stays put, and the SessionEnd hook emits a visible `systemMessage` telling the user compression failed and why. `apsolut-cortex status` now shows circuit-breaker state (recent failures + last-failure age).
- **BM25 search matches non-adjacent terms.** The FTS5 query wrapped the entire search string as a single quoted phrase, so any multi-word query whose words weren't contiguous in a memory returned zero BM25 results (e.g. "libsql credential manager" missed "libsql encryption key … Credential Manager"). Terms are now quoted individually and OR-joined; injection safety (quote doubling) is preserved.
- **`db re-encrypt` no longer drops tables.** The copy list was hardcoded to the five Phase-1 tables and silently dropped `raw_messages` (all raw transcripts) and `memory_tags` during re-encryption. The table list is now derived from the source database's `sqlite_master` (FTS shadow tables excluded — triggers repopulate them), so future migration-added tables survive automatically. An unknown-table failure now aborts before the swap with the source DB untouched.
- **7 strict-mode TypeScript errors** in cli.ts, embed.ts, session-start.ts, and mcp/server.ts. `tsc --noEmit` is now enforced via a `typecheck` script wired into `prepublishOnly` and a new GitHub Actions CI workflow (typecheck + tests on push/PR).
- **Windows: libSQL file handles are released deterministically** during backup/re-encrypt file swaps (Bun frees the handle only at GC; the retry loops now nudge GC). Also cuts the backup test suite from ~9s to under 1s.

### Known limitations
- **M3 encryption does not work on native Linux** — libSQL's local `encryptionKey` mode fails with `SQLITE_IOERR` (encrypted DB cannot be read back). Discovered by the first CI run on ubuntu-latest; the re-encrypt test suite is skipped on Linux and the README stability note now says so. Windows and macOS unaffected. `db re-encrypt` now refuses on Linux with a clear message instead of producing an unreadable database.

## [0.12.6] – 2026-05-25

### Changed
- Version-only bump. No functional or doc changes vs. 0.12.5; this release exists so the published npm package can be refreshed. Tarball contents are byte-identical apart from `package.json`'s `version` field and this CHANGELOG entry.

## [0.12.5] – 2026-05-25

### Added
- **`LICENSE`** at repo root. `package.json` already declared MIT but the actual file was missing; npm shows the badge from the field but legal/auditors want the file. Now both line up.
- **`SECURITY.md`** — supported-versions table, what to report vs. what's out of scope, private disclosure channels (GitHub security advisories or maintainer email).
- **`CONTRIBUTING.md`** — setup, code conventions (no AI attribution lines, no future-proofing comments, the `~/.apsolut/` rule), test patterns, migration safety rules, ADR convention.
- **`CODE_OF_CONDUCT.md`** — Contributor Covenant v2.1 reference.

### Changed (docs only)
- **README `## Configuration`** trimmed from ~21 inline env vars to a "five most commonly tweaked" table + a pointer to the full reference in `docs/CONFIG.md`. Removed the duplication that was getting out of sync between the two files.

## [0.12.4] – 2026-05-25

### Changed (docs only)
- **README** — added a new "Stability notes" section after the roadmap, flagging M3 encryption as **work-in-progress / experimental** (the keychain backend is platform-fragile on WSL2 and a file-based key fallback is not yet shipped; key loss = data loss). M3 roadmap line demoted from `[x]` to `[~]` with an inline ⚠ marker pointing at the new section. Same section also flags M6 as new + opt-in, the sparse eval signal, and the deferred M7/M8.
- Renamed sibling-project link from `apsolut-scaffolding` to **`apsolut-seshat`** (project was renamed upstream 2026-05-25).

## [0.12.3] – 2026-05-25

### Changed (docs only)
- **README.md** — `## Commands` rewritten to list every shipped subcommand grouped by purpose (setup / daily / curation / ops / eval). MCP tools table gained `memory_recall`. "How it works" gained the M6 in-session + reflector + auto-export bullets. Storage layout shows `logs/`, `obsidian/`, `buffer/`, `backup/`. Configuration section lists the M6 (`OBSERVE_THRESHOLD`, `OBSERVE_BLOCK_MULT`, `REFLECT_THRESHOLD`), M4 (`RAW_RETENTION_DAYS`), and M1 (`SHADOW`) env vars.
- **docs/CONFIG.md** — full env var reference rebuilt; ~21 vars grouped by concern with defaults inline.
- **docs/OPERATIONS.md** — added `## Hook setup` (legacy vs M6 `install-hooks`), `## Curation` (M5 promote/demote/tag/grep/delete/correct), `## Vault export` (M5 obsidian flow).
- **docs/STORAGE.md** — directory layout now shows every dir cortex writes to; schema section lists every table with origin migration and purpose (added `raw_messages`, `memory_tags`, `_migrations`, `_migrations_lock`).
- **docs/PHASE2-BUILD-ORDER.md** — added a status table at the top showing M0–M6 done and which version each shipped in, M7/M8 deferred with reasons.

## [0.12.2] – 2026-05-25

### Security
- Bumped `@anthropic-ai/sdk` from `^0.88.0` to `^0.91.1` to clear a Dependabot advisory. Our only usage (`client.messages.create()` in `src/compress.ts`) is unchanged between these versions; 72/72 tests still pass.

## [0.12.1] – 2026-05-25

### Fixed
- **Tarball size:** `gpt-tokenizer` (added in 0.11.0 for M6 token counting) was not in `build.ts`'s external list and got bundled into every hook script — pushed the published tarball from ~30 KB to **4.2 MB**. Added to externals; tarball back down to ~73 KB.

## [0.12.0] – 2026-05-25

### Added
- **M5 visibility layer — finished (Phase 2):**
  - **Compiled views** in the Obsidian vault export: `by-category/<cat>.md` (every memory of a category across all projects, sorted by weight), `by-project/<name>.md` (per project, grouped by category), `_health.md` (low-trust high-use → consider `promote`; high-trust stale > 60d → consider `demote`; flagged-by-`memory_contradict`). All compiled views regenerate from scratch on every full export and stale files are GC'd.
  - **Curation CLI commands:**
    - `apsolut-cortex promote <id>` / `demote <id>` — walk the trust ladder (`observed → validated → proven → canonical`).
    - `apsolut-cortex tag <id> <tag>` / `untag <id> <tag>` — free-form labels. Tags are lowercased and de-duped; `INSERT OR IGNORE` on the composite PK.
    - `apsolut-cortex grep <pattern>` — substring search across content + context in the current project (top 50 hits).
    - `apsolut-cortex delete` with `--id`, `--project`, `--tag`, `--before YYYY-MM-DD`, `--grep <pat>`. Filters combine with AND. Always shows a preview of the first 5 matches; refuses to run without `--yes`. No raw SQL accepted by design.
  - **Tags in frontmatter:** memory `.md` files now include `tags:` array and a `[[tag-<name>]]` link footer for Obsidian graph view.
- **Migration 004 — `memory_tags` table:** `(memory_id, tag, created_at)` with composite PK and a `(tag, memory_id)` index. Append-only; deletion is GC'd when the parent memory is removed.
- **`src/curation.ts`** module with `promoteMemory` / `demoteMemory` / `tagMemory` / `untagMemory` / `getTagsForMemory` / `grepMemories` / `previewDeletion` / `applyDeletion` — 12 unit tests covering the trust ladder, tag round-trip, all delete-filter shapes, and the invalid-date guard.

## [0.11.0] – 2026-05-25

### Added
- **M6 in-session compression (opt-in, Phase 2):** memories now land *during* the session, not just at SessionEnd. Triggered by conversation token budget (default 30000) with a hard safety net at 1.2× that threshold. Activates the M4 `raw_messages` table — `memory_recall(id)` now returns actual transcript history for any memory produced after upgrade.
  - **`src/tokens.ts`** — `gpt-tokenizer`-backed local estimator. Never calls a model just to count tokens. Handles structured-block content (`text`, `tool_use`, `tool_result`) flattening for Claude Code transcript format.
  - **`src/buffer.ts`** — per-session JSONL spill + single-flight lockfile + cursor (last-compressed `msg_idx`). Survives process kills; cursor makes re-runs idempotent.
  - **`src/transcript.ts`** — `readTranscript()` / `sliceRange()` / `persistRawMessages()` / `captureTranscript()`. Parses the JSON-l transcript Claude Code hands every hook via `transcript_path`.
  - **`src/compress-runner.ts`** — shared `compressSlice()` used by PreCompact, the detached worker, and SessionEnd. Persists raw messages, slices by cursor, compresses via existing `compress.ts` pipeline (Haiku → Ollama fallback + circuit breaker), inserts memories with `source_session_id` / `source_start_msg_idx` / `source_end_msg_idx` set so `memory_recall` can resolve them later.
  - **`src/hooks/pre-compact.ts`** — synchronous emergency capture before Claude Code compacts. Waits up to 3 s for any running worker, then force-acquires the lock.
  - **`src/hooks/compress-worker.ts`** — detached background worker spawned by PostToolUse when the token budget is hit. Honors the single-flight lock and exits quietly if another worker is already running.
  - **`src/hooks/post-tool-use.ts`** — updated with the token-budget trigger; spawns the worker async at threshold, falls back to synchronous compression at 1.2× threshold.
  - **`src/hooks/session-end.ts`** — calls `compressSlice` for the final tail of the session so every memory has a source range, then `clearAllForSession` cleans up buffer artifacts.
  - **`src/reflector.ts` + SessionEnd integration** — when this session's memories exceed `APSOLUT_CORTEX_REFLECT_THRESHOLD` tokens (default 40000), re-summarize them into denser `tier=meta`, `source=reflector` memories. Conservative v1: SessionEnd only.
  - **`templates/hooks-m6.json`** + **`apsolut-cortex install-hooks`** CLI — opt-in hook installer that wires `SessionStart + PostToolUse + Stop + SessionEnd + PreCompact` into `~/.claude/settings.json`. Existing users on legacy `init`-installed hooks are not affected until they explicitly opt in.
- **3 new env vars in `src/config.ts`:** `APSOLUT_CORTEX_OBSERVE_THRESHOLD` (30000), `APSOLUT_CORTEX_OBSERVE_BLOCK_MULT` (1.2), `APSOLUT_CORTEX_REFLECT_THRESHOLD` (40000).
- **3 new test files (tokens, buffer, transcript)** — 25 new tests; 59/59 total tests passing across 10 files.
- **`docs/decisions/002-async-compression.md`** explains why we use detached child processes instead of the speculated `"async": true` hook flag (which doesn't exist in Claude Code 2.x).
- `CHANGELOG.md` and `templates/` added to the published `files` array in `package.json`.

### Dependencies
- Added `gpt-tokenizer@^3.4.0` for local conversation-token counting.

## [0.10.0] – 2026-05-24

### Added
- **M5 visibility layer (partial, Phase 2):** `apsolut-cortex export` writes one markdown file per memory to `~/.apsolut-cortex/obsidian/memories/` with full YAML frontmatter (id, project, tier, category, trust, weight, used_count, created_at, last_used, source, source_session_id). Body includes the memory content, optional context section, and `[[wiki-links]]` to the project + category. The vault is regenerated from scratch each time; orphaned `.md` files (memories no longer in the DB) are garbage-collected.
- **`index.md`** at the vault root: grouped by project → category, sorted by weight within each group, with `[[link]]`s to every memory file. The header marks the vault as generated output.
- **Auto-export on `SessionEnd`:** the existing session-end hook now calls `exportVault` after compression + decay/prune. Wrapped in its own try/catch so a markdown write failure can't break compression.
- **`apsolut-cortex export`** CLI command for manual triggers.
- 4 export tests against an isolated tmp vault dir — file shape, frontmatter, orphan GC, project-filter mode. 34/34 total tests passing.

### Pending in M5 (rolled into a future commit, not blocking 0.10.0)
- `apsolut-cortex promote` / `demote` / `tag` / `grep` / `delete` CLI commands.
- Compiled views: per-category pages, per-project pages, `_health.md` for low-trust / contradiction-candidate memories.
- Inter-memory `[[wiki-links]]` based on content overlap.

## [0.9.0] – 2026-05-24

### Added
- **M4 range-linked memories (Phase 2):** two new migrations and a new MCP tool let compressed memories point back at the raw conversation slice they were derived from.
  - Migration 002 adds nullable `source_session_id`, `source_start_msg_idx`, `source_end_msg_idx` columns to `memories` + a `(source_session_id, source_start_msg_idx)` index. PRAGMA check before each ADD COLUMN so re-runs are no-ops.
  - Migration 003 creates `raw_messages(session_id, msg_idx, role, content, created_at)` with composite primary key + `(session_id, created_at)` index. Append-only.
  - `insertMemory()` accepts optional `source_session_id` / `source_start_msg_idx` / `source_end_msg_idx`. Existing call sites unchanged (defaults to NULL).
  - `insertRawMessage()`, `getRawRange()`, `getMemoryWithRange()` helpers added to `src/db.ts`.
  - New MCP tool `memory_recall(id)` — returns the raw transcript slice for a memory, or a clear message if the memory predates M4 (NULL ranges) or the raw window was pruned by retention.
- `APSOLUT_CORTEX_RAW_RETENTION_DAYS` env var (default 90) documented in `src/config.ts`. The cleanup job lands with M8's `is_pinned` work; until then all raw_messages rows are retained.

## [0.8.0] – 2026-05-24

### Added
- **M3 encryption at rest + backup/restore (opt-in, Phase 2):** libSQL-native encryption via `encryptionKey` on `createClient()`. Key is stored in the OS keychain (`@napi-rs/keyring` → Windows Credential Manager / macOS Keychain / libsecret) under service `apsolut-cortex`, account `db-encryption-key`.
- **`apsolut-cortex db re-encrypt`** CLI command: opt-in migration of the existing plaintext DB. Generates a 256-bit key, snapshots the current DB to `~/.apsolut-cortex/backup/pre-encrypt-<ts>.db` (never deleted), copies every row into a fresh encrypted DB, atomically replaces the live file. Dry-run by default; requires `--yes` to proceed. Original DB untouched on any failure.
- **`apsolut-cortex backup`** CLI command: physical file snapshot to `~/.apsolut-cortex/backup/manual-<ts>.db`.
- **`apsolut-cortex restore [<path>] [--yes]`** CLI command: lists snapshots without args; restores with `--yes`. Always writes a pre-restore safety snapshot first.
- **`src/keyring.ts`** + **`src/backup.ts`** modules with focused unit tests. Round-trip test verifies that a re-encrypted DB reads back correctly with the key, fails to open without it, and that FTS5 search still works after re-encryption.
- **`docs/OPERATIONS.md`** updated with the encryption opt-in flow, backup/restore commands, and the "do not nuke the keychain after encrypting" warning.

### Changed
- `getDb()` now checks the OS keychain on first connection. If a key is present it is passed to `createClient({ encryptionKey })`; if absent, the DB opens unencrypted (same behavior as before this release). Existing users are unaffected until they explicitly run `apsolut-cortex db re-encrypt --yes`.
- `build.ts` marks `@napi-rs/keyring` as external so its native bindings resolve from `node_modules/` at runtime rather than being bundled.

### Dependencies
- Added `@napi-rs/keyring@^1.3.0` for OS keychain access (native, ~few hundred KB per platform).

## [0.7.0] – 2026-05-24

### Added
- **M2 retrieval audit log (Phase 2):** every `memory_search` call now appends one JSONL line to `~/.apsolut-cortex/logs/retrievals.jsonl` with the query, per-candidate scores (BM25 rank, vector rank, final rank), tier/trust/weight, what was actually injected back to Claude, and end-to-end latency. Shadow mode entries continue to route to `shadow.jsonl` so the production log stays clean.
- **`apsolut-cortex correct`** CLI command: reads the most recent retrieval and flags it as a miss in `~/.apsolut-cortex/logs/corrections.jsonl`. Pass `--with "<correct answer>"` to also insert the correction as a new `correction`-category memory linked back to the failing retrieval. Karpathy's "outputs feed back in" — the user's miss-labeling gesture now grows the knowledge base instead of just instrumenting it.
- **`src/logs.ts`** module: shared JSONL append + read helpers, used by the MCP server, the CLI `correct` command, and tests. Refactors the ad-hoc shadow-logging in `mcp/server.ts` so all three log streams share one implementation.

### Changed
- `hybridSearch()` in the MCP server now returns per-source ranks alongside the result list so the audit log can record where each memory came from in the BM25 vs vector lists.

## [0.6.2] – 2026-05-24

### Fixed
- `apsolut-cortex eval` now prints a clean "maintainer-only" message when invoked from a global npm install, instead of dying on a missing `evals/runner.ts` (the eval harness lives outside the published tarball by design). Help banner cleaned up accordingly.

## [0.6.1] – 2026-05-24

### Changed
- README's storage section corrected: the on-disk format is libSQL (Turso's SQLite fork), not vanilla SQLite. Added a short note that any `sqlite3` CLI can still read the file and that migration to Turso cloud is a URL change.
- `docs/STORAGE.md` expanded with the three concrete `createClient` shapes for the future Turso migration path (local file / pure cloud / cloud + embedded replica). All work identically against the existing schema — preserves the option without committing to it.

## [0.6.0] – 2026-05-24

### Added
- **M1 eval harness (Phase 2):** `evals/` directory with `golden.jsonl` (5 seeded entries, target 30), `evals/fixtures/seed.ts` (reproducible in-memory fixture DB with 10 known memories), `evals/runner.ts` (hit rate + MRR computation), `apsolut-cortex eval run` and `apsolut-cortex eval baseline` CLI subcommands. Each run scores **two** retrievals side-by-side: the production hybrid path and a tokenized-grep baseline. The runner prints a verdict line — if hybrid beats grep by ≥5pp we keep the complexity; if grep matches or wins we have license to simplify in M8. At 5 seeded queries with hand-picked content both score 5/5; the gap will only appear at 20+ entries with more paraphrased queries (per the eval README guidance).
- **`searchGrep(db, projectId, query, limit)`** in `src/db.ts` — Karpathy "LLM reads the markdown" baseline. Tokenizes the query (≥3 chars, stop-words dropped), scores by token-overlap count, breaks ties by recency.
- **Shadow mode** in the MCP server: when `APSOLUT_CORTEX_SHADOW=true`, `memory_search` still runs retrieval but returns nothing to Claude. Would-be matches are appended to `~/.apsolut-cortex/logs/shadow.jsonl` for offline analysis. Lets us tune retrieval against real sessions without polluting the conversation.

## [0.5.7] – 2026-05-24

### Added
- **M0 pre-flight (Phase 2):** `bun:test` smoke test, migration system (`src/migrations/`, `_migrations` table, runner with sentinel-row lock), `apsolut-cortex migrate` CLI command, `CHANGELOG.md`, `docs/` scaffolding (`OPERATIONS.md`, `STORAGE.md`, `PROVIDERS.md`, `CONFIG.md`, `OLLAMA.md`, `decisions/`).
- **001-initial-schema migration:** existing schema extracted into `src/migrations/001-initial-schema.ts`. Runner detects existing DBs and back-fills the migration row without re-running SQL.

### Changed
- `~/.apsolut/` is now documented as a permanent invariant — never to be used by this plugin. The path is reserved for other `apsolut-*` tools. `.npmignore`, `README.md`, and `docs/PHASE2-BUILD-ORDER.md` updated to frame this as a rule, not a one-time rename.

### Roadmap
- Karpathy's personal-LLM-wiki insights folded into the roadmap: M1 gains a `searchGrep` baseline (test whether hybrid RAG actually beats substring match at our scale), M2 gains bidirectional capture (`apsolut-cortex correct --with "..."`), M5 gains compiled views (`index.md`, per-category pages, `_health.md`, inter-memory `[[wiki-links]]`), and M8 gains a standing `apsolut-cortex audit` linting command.

## [0.5.6] – 2026-04-21

### Fixed
- Skill docs and onboarding banner aligned with the real MCP tool schemas.
- Skill names made consistent across onboarding.
- `hono` pinned to 4.12.14 (GHSA-458j-xx4x-4375).
- Replaced `@xenova/transformers` with `@huggingface/transformers` (CVE-2026-41242).

## [0.5.3] – 2026-04-19

### Fixed
- Skill names renamed to `apsolut-*` prefix to avoid colliding with Claude built-in skills.

### Added
- `[apsolut-cortex]` prefix on all output.

## [0.5.2] – 2026-04-19

### Added
- `repository`, `homepage`, and `bugs` fields in `package.json` for npm metadata.

## [0.5.1] – 2026-04-19

### Changed
- README env-var table rewritten as a list for better readability on npm.

## [0.5.0] – 2026-04-19

### Changed
- Env var prefix renamed from `CORTEX_*` to `APSOLUT_CORTEX_*`.

### Fixed
- README clarified: sessions are automatic, not a clean-start ritual.

## [0.4.1] – 2026-04-13

### Changed
- Data namespace renamed from `~/.apsolut/` to `~/.apsolut-cortex/` (reserves `~/.apsolut/` for other `apsolut-*` tools).
- Phase 2 roadmap added to README; install instructions and storage path corrected.
- `.mcp.json` gitignored (machine-specific paths).
- Phase 2 build order plan added (v3, fact-checked against the codebase).

### Fixed
- Memory system hardening: pruning bug, transaction safety, atomic writes, config externalization.

## [0.4.0] – 2026-04-12

### Added
- Initial public release on npm.

[Unreleased]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.11...HEAD
[0.12.11]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.10...v0.12.11
[0.12.10]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.8...v0.12.10
[0.12.8]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.6...v0.12.8
[0.12.6]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.5...v0.12.6
[0.12.5]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.4...v0.12.5
[0.12.4]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.3...v0.12.4
[0.12.3]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.2...v0.12.3
[0.12.2]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.1...v0.12.2
[0.12.1]: https://github.com/apsolut/apsolut-cortex/compare/v0.12.0...v0.12.1
[0.12.0]: https://github.com/apsolut/apsolut-cortex/compare/v0.11.0...v0.12.0
[0.11.0]: https://github.com/apsolut/apsolut-cortex/compare/v0.10.0...v0.11.0
[0.10.0]: https://github.com/apsolut/apsolut-cortex/compare/v0.9.0...v0.10.0
[0.9.0]: https://github.com/apsolut/apsolut-cortex/compare/v0.8.0...v0.9.0
[0.8.0]: https://github.com/apsolut/apsolut-cortex/compare/v0.7.0...v0.8.0
[0.7.0]: https://github.com/apsolut/apsolut-cortex/compare/v0.6.2...v0.7.0
[0.6.2]: https://github.com/apsolut/apsolut-cortex/compare/v0.6.1...v0.6.2
[0.6.1]: https://github.com/apsolut/apsolut-cortex/compare/v0.6.0...v0.6.1
[0.6.0]: https://github.com/apsolut/apsolut-cortex/compare/v0.5.7...v0.6.0
[0.5.7]: https://github.com/apsolut/apsolut-cortex/compare/v0.5.6...v0.5.7
[0.5.6]: https://github.com/apsolut/apsolut-cortex/compare/v0.5.3...v0.5.6
[0.5.3]: https://github.com/apsolut/apsolut-cortex/compare/v0.5.2...v0.5.3
[0.5.2]: https://github.com/apsolut/apsolut-cortex/compare/v0.5.1...v0.5.2
[0.5.1]: https://github.com/apsolut/apsolut-cortex/compare/v0.5.0...v0.5.1
[0.5.0]: https://github.com/apsolut/apsolut-cortex/compare/v0.4.1...v0.5.0
[0.4.1]: https://github.com/apsolut/apsolut-cortex/compare/v0.4.0...v0.4.1
[0.4.0]: https://github.com/apsolut/apsolut-cortex/releases/tag/v0.4.0