cortex-mcp 2.7.6 → 2.9.0

package/CHANGELOG.md

@@ -1,58 +1,305 @@
-# Changelog
-
-All notable changes to `cortex-mcp` are documented here.
-
----
-
-## [1.0.6] — 2026-03-14
-
-### Everything Is Now 100% Free
-- Removed all feature gates, paywalls, and license checks
-- All 35 tools available to every user with no restrictions
-- Limits auto-reset every 24 hours (no more "restart to reset")
-
-### Critical Bug Fixes
-- **FIXED:** `memory-store.js` — `this` binding lost in `.map()` callbacks caused `getActive()`, `getByType()`, `getByFile()` to crash in strict mode
-- **FIXED:** `git-hooks.js` — Wrong database path (`cortex.db` → `data/cognitive.db`) caused all branch-switch and merge memories to be silently dropped
-- **FIXED:** `git-hooks.js` — Missing `created_at` NOT NULL column in INSERT statement
-- **FIXED:** `git-capture.js` — Same missing `created_at` column bug
-- **FIXED:** `clear_memories` — Crashed with JS spread operator `...params` in better-sqlite3 `.all()`
-- **FIXED:** `search_timeline` — Same spread crash
-- **FIXED:** `backup_brain` — Pointed to wrong property for DB file path (`dbPath` → `memoryStore.db.name`)
-- **FIXED:** `undo_last` — Was ordering by `timestamp` (last-accessed) instead of `created_at` (birth time), causing wrong memory to be undone
-
-### New Tools (3 Added)
-- **`export_memories`** — Export entire brain to a shareable JSON file
-- **`import_memories`** — Import memories from JSON (skips duplicates)
-- **`get_related_memories`** — Graph traversal: find memories connected by relationship edges
-
-### New Features
-- **Dashboard fully interactive** — Delete and pin memories directly from `localhost:3456`
-- **Dashboard live refresh** — Auto-refreshes every 30 seconds
-- **Dashboard dark theme** — Beautiful, professional UI with Inter font
-- **24h auto rate-limit reset** — Limits reset automatically, no server restart needed
-- **MEMORY_CAP raised** — Maintenance (decay/consolidation) now spans up to 10,000 memories
-- **MAX_CONTEXT_CHARS raised** — Context injected per request raised from 3,000 to 12,000 characters
-- **Context configurable** — Set `CORTEX_MAX_CONTEXT_CHARS` env var to customize
-- **Memory IDs in recall output** — `recall_memory` now shows IDs so you can immediately pin/downvote
-- **Correct tool names in setup rules** — AI instruction files now reference `force_recall` (not `mcp_cortex_force_recall`)
-
-### Performance
-- Added `idx_signals_memory` index on `user_signals(memory_id)` — 100x faster feedback lookups
-- All database maintenance operations raised from 500 → 10,000 memory cap
-
----
-
-## [1.0.4] — 2026-02-xx
-
-### Initial public release
-- Core MCP server with stdio transport
-- 25 tools for memory storage, recall, and project scanning
-- SQLite database with FTS5 full-text search
-- Vector search with `@xenova/transformers` (local, no API key)
-- Auto-learning from AI responses
-- Memory decay and confidence scoring
-- Memory consolidation (dedup)
-- Git commit capture via hooks
-- Web dashboard at localhost:3456
-- Auto-setup for Cursor, Claude, Windsurf, Copilot, Zed, Antigravity
+# Changelog
+
+All notable changes to Cortex MCP Server are documented here.
+
+## [2.6.0] -- 2026-03-04
+
+### Added — OpenClaw-Style Autonomous Brain
+
+- **SOUL.md** — Persistent identity file at `.cortex/SOUL.md` storing user preferences and conventions. Auto-learns from high-access memories via `autoLearnSoul()`.
+- **Daily Diary** — Timestamped Markdown logs at `~/.cortex-memory/diary/YYYY-MM-DD.md`. Entries written on every `store_memory` and `auto_learn`. Shown as Layer 7.5 in `force_recall`.
+- **MEMORY.md** — Human-readable knowledge export at `.cortex/MEMORY.md`. Auto-exported after `auto_learn` stores ≥2 memories. Auto-imported on startup from user edits.
+- **MMR Re-ranking** — Maximal Marginal Relevance (Carbonell & Goldstein 1998) applied to all recall results. λ=0.7 (70% relevance, 30% diversity). Uses Jaccard similarity on tokenized content.
+- **Embedding Cache** — LRU cache (256 entries) for vector embeddings. Skips re-embedding identical queries. Reduces `recall_memory` latency by ~50%.
+- **Query Expansion** — 40+ synonym groups (auth→authentication/login/jwt, db→database/sql/postgres, etc.). Generates alternative queries when FTS returns 0 results.
+- **Access Pattern Tracking** — Records which memory types are recalled most. Personalizes ranking with 0.7x–1.8x boost multipliers. Profile loaded on startup via `loadAccessProfile()`.
+- **Cross-Memory Linking** — `autoLinkMemory()` connects new memories to existing ones by shared files, tags, and word overlap.
+- **Two-Tier Decay** — Transactional memories (INSIGHT) decay at 5 days. Operational memories (DECISION, CORRECTION, BUG_FIX, CONVENTION) protected for 60 days.
+
+### Changed
+
+- **Brain pipeline** — Expanded from 12 to 14 layers (added Soul, Day 1 scan, Diary layers)
+- **`expandQuery`** — Replaced basic 10-synonym map with OpenClaw-style 40+ synonym module
+- **`handleHealthCheck`** — Now shows diary stats, access profile status, and top memory types
+- **`handleRecallMemory`** — Added embedding cache, MMR re-ranking, alternative query fallback
+
+### Fixed
+
+- **Diary logging gap** — `store_memory` now writes diary entries (previously only `auto_learn` did)
+- **2 lint warnings** — `profileMemories` and `MemoryType` prefixed with underscore
+
+## [2.4.0] -- 2026-03-03
+
+### Added
+
+- **Independent importance scoring** — Memories scored by type (CORRECTION=0.85, INSIGHT=0.55) + content signal boosts (file paths, error keywords, version numbers)
+- **Topic tag extraction** — Auto-tags memories with technologies, domains, and actions (`['typescript', 'auth', 'database', 'testing']`)
+- **FTS resolved filter** — Full-text search excludes resolved-tagged memories from results
+- **Code block safety** — `splitSentences` strips ``` fenced blocks and inline code before extraction
+- **Dashboard auto-refresh** — Auto-refreshes every 30 seconds via meta tag
+
+### Changed
+
+- **Action field** — Now stores source (`auto_learn:decision`) instead of duplicating intent content
+
+## [2.3.1] -- 2026-03-03
+
+### Removed
+
+- **Dead code: `event-bus.ts`** — 60 lines, 0 imports anywhere in codebase
+- **Dead method: `cleanFTSQuery()`** — Unused private method in hybrid-retriever.ts
+
+### Added
+
+- **Memory decay Step 5** — Deactivate resolved memories >7 days old
+
+### Fixed
+
+- **Lint** — Prefixed unused `_confidence` parameter in completion-resolver.ts
+- **README** — Updated "What's New in v2.0" → "What's New in v2.3"
+
+## [2.3.0] -- 2026-03-03
+
+### Fixed (Critical Brain Bugs)
+
+- **Markdown bullet extraction** — `splitSentences()` now splits on newlines and bullets, not just periods. 4x more memories extracted from AI responses
+- **Solo developer capture** — Added "I used/chose/picked" patterns (was only "we used")
+- **DECISION vs CORRECTION misclassification** — Raised DECISION confidence to 0.78 (was 0.70, CORRECTION's "instead of" won at 0.75)
+- **Completion resolver wired** — `completion-resolver.ts` was dead code, now integrated into `handleAutoLearn`
+- **Ranker resolved penalty** — Resolved-tagged memories score 85% lower (0.15x multiplier)
+- **Context builder filter** — Excludes resolved memories from context output
+
+## [2.2.0] -- 2026-03-02
+
+### Added
+
+- **20 MCP tools** — Added `review_code`, `pre_check`, `check_impact`, `resume_work`
+- **Attention ranking** — Debugging context boosts bug-fix memories, coding boosts conventions
+- **Meta-memory** — Knowledge gap detection for files with zero memories
+- **Memory consolidation** — Duplicate memories merge, keeping highest importance
+
+### Fixed
+
+- **Git hooks cross-platform** — Changed `cortex-capture` → `npx cortex-capture` in hook scripts for reliable binary resolution
+- **Windows chmod** — Separated `writeFileSync` from `chmodSync` with try/catch (Windows ignores chmod safely)
+- **Lint cleanup** — Removed unused `getTipOfTheDay` import from mcp-handler.ts
+
+## [2.0.0] -- 2026-03-02
+
+### Fixed
+
+- **Windows git command** — `resume-work.ts` used `2>/dev/null` (Unix-only), changed to `stdio: ['pipe','pipe','pipe']`
+
+### Changed
+
+- **Setup CLI** — `getServerEntryWithEnv()` auto-passes `OPENROUTER_API_KEY` from system environment
+- **Setup CLI** — Prints free LLM tip with OpenRouter instructions after setup
+- **Setup CLI** — Added `transportType: 'stdio'` to server entry for proper MCP protocol compliance
+- **get_stats** — Returns 6 new fields: `streak`, `longestStreak`, `totalAutoLearns`, `successPatternsLearned`, `errorsLearned`, `llmProvider`
+- **README** — "What's New" section updated from v1.4 to v2.0
+
+## [1.9.0] -- 2026-03-02
+
+### Added
+
+- **Daily streak tracking** — Tracks consecutive days of Cortex usage with "Don't break the chain!" gamification
+- **Streak display** — Shows "🔥 7-day streak!" in stats footer (3-day, 7-day, 14-day, 30-day tiers)
+- **Tip of the Day** — Surfaces a random useful memory in force_recall footer (deterministic daily pick)
+- **Success tracker wired** — `auto_learn` now detects praise ("perfect", "that worked") and stores proven approaches
+- **Error learner wired** — `auto_learn` now captures error fingerprints (TS errors, stack traces, npm failures)
+- **trackSuccess()** / **trackErrorLearned()** — New usage-stats tracking functions
+
+## [1.8.0] -- 2026-03-02
+
+### Added
+
+- **Project memory filtering** — Memories tagged with current project get 1.3x score boost, other projects get 0.7x penalty
+- **Error fingerprint patterns** — 7 new signals in auto-learner: TypeError, ENOENT, HTTP 4xx/5xx, npm ERR, stack traces, build failures
+- **Git-enhanced resume** — `resume_work` shows recent git commits from last 24 hours
+
+## [1.7.0] -- 2026-03-02
+
+### Added
+
+- **OpenRouter free LLM** — Add `OPENROUTER_API_KEY` for smarter memory extraction at zero cost
+- **Default free models** — `meta-llama/llama-4-scout:free` and `deepseek/deepseek-chat-v3-0324:free`
+- **OpenRouter headers** — Proper `HTTP-Referer` and `X-Title` headers
+- **15-second timeout** — Extended from 10s for slower free tier models
+
+## [1.6.0] -- 2026-03-02
+
+### Added
+
+- **Brain Health Score** — Composite 0-100 score based on usage patterns (Newborn → Genius grades)
+- **Saved You counter** — Tracks times Cortex prevented repeated mistakes
+- **Memory milestones** — Celebrations at 10, 25, 50, 100, 250, 500 memories
+- **Time saved estimate** — Estimates hours saved based on recalls and catches
+- **Stats footer** — Every `force_recall` response ends with stats summary
+
+## [1.5.0] -- 2026-03-02
+
+### Added
+
+- **5 new auto-learn patterns** — Enhanced pattern matching for decisions, conventions, insights
+- **Topic-aware force_recall** — Search memories relevant to the current topic
+- **Project auto-detection** — Detects project name from package.json or directory name
+
+### Changed
+
+- **Stronger tool descriptions** — `force_recall`, `auto_learn`, `pre_check` now include consequence language so AI uses them consistently
+- **Raised text limit** — `auto_learn` content limit increased from 5K to 50K chars (AI responses can be very long)
+- **Rule templates** — Ships rules for Claude (`CLAUDE.md`), Cursor (`.cursorrules`), Windsurf (`.windsurfrules`), Copilot (`.github/copilot-instructions.md`), and Gemini (`GEMINI.md`)
+- **Updated README** — Reflects v1.4 features
+
+## [1.4.1] -- 2026-03-02
+
+### Fixed
+
+- **Pattern priority bug** — `extractMemories` now picks highest-confidence match, not first match
+- **CONVENTION patterns** — Split into specific (0.85 confidence) and generic (0.65) groups. "best practice" now correctly classified
+- **INSIGHT patterns** — Split into specific (0.85) and generic (0.55). "for future reference" now correctly classified
+- **Preference regex** — Fixed "be more concise" detection
+
+## [1.4.0] -- 2026-03-02
+
+### Added
+
+- **instructions-generator** — Converts memories into DO/DON'T/WATCH-OUT/STYLE rules
+- **tool-recommender** — Recommends Cortex tools based on context (🔴 must / 🟡 should / 🟢 could)
+- **regression-guard** — Extracts verification steps and attaches to BUG_FIX memories
+- **error-learner** — Learns from build/compile error patterns
+- **success-tracker** — Reinforces approaches when user says "perfect" or "works"
+- **file-relationships** — Tracks which files are typically co-edited
+- **preference-learner** — Detects user communication and coding preferences
+- **32 capture patterns** — Enhanced from 14 to 32 in auto-learner
+
+### Changed
+
+- **Lint cleanup** — Fixed 10 unused import warnings in mcp-handler.ts
+- **Branding** — Server name is now "Cortex" (capital C) for proper IDE display
+
+## [1.3.0] -- 2026-03-01
+
+### Added
+
+- **Architecture graph** — Maps file roles and dependencies
+- **Export map** — Tracks what functions are exported from which files
+- **Convention detector** — Auto-detects coding conventions during project scan
+- **Knowledge gap reporting** — Identifies uncovered areas in scan results
+- **Architecture-aware pre_check** — Shows file role and dependencies
+
+## [1.2.1] -- 2026-02-24
+
+### Changed
+
+- **Shared tag extraction** — Created `src/utils/extract-tags.ts` as single source of truth for topic-based tag extraction
+- **Eliminated duplicate code** — `llm-enhancer.ts` and `git-memory.ts` now import from the shared utility
+
+## [1.2.0] -- 2026-02-24
+
+### Performance (Critical)
+
+- **`force_recall` parallelization** — Brain layers 6-12 now run concurrently via `Promise.allSettled()`. Latency drops from ~15-19s to ~3-5s
+- **`force_recall` caching** — Results cached with 30s TTL, repeated calls return instantly
+- **Removed double FTS sanitization** — `hybrid-retriever.ts` no longer double-sanitizes queries
+- **Fixed sync `touch()` in `Promise.all`** — Replaced with `runTransaction()` for proper batching
+
+### Performance (Database)
+
+- **Context builder** — Consolidated 7 SQL queries into 1 (`context-builder.ts`)
+- **Event log** — Cached 4 prepared statements in constructor (`event-log.ts`)
+- **Anticipation engine** — FTS search instead of load+filter (`anticipation-engine.ts`)
+- **Session tracker** — `findByTag()` instead of JS filter (`session-tracker.ts`)
+- **Meta-memory** — Reduced fetch from 500 to 200 (`meta-memory.ts`)
+
+### Performance (Memory)
+
+- **Hoisted stopwords** — Module-level constants in `learning-rate.ts` and `memory-quality.ts`
+- **Proactive cache eviction** — Stale entries swept during `setCache()` (`memory-cache.ts`)
+
+### Robustness
+
+- **Session array capping** — Arrays capped at 20 entries to prevent unbounded growth
+- **Contradiction detection** — Raised overlap threshold from 2 to 3 (fewer false positives)
+- **Top-level imports** — Moved `require('child_process')` to top-level in `git-memory.ts`, `temporal-engine.ts`
+- **Export-import** — Dedup cap reduced from 5000 to 1000
+
+### Architectural
+
+- **Tag-based scan detection** — Replaced fragile magic-string check with `findByTag('project-scan')` in `project-scanner.ts`
+
+## [1.1.1] -- 2026-02-23
+
+### Fixed
+
+- **Memory store** — Cached prepared statements for all read queries (Round 4 optimization)
+- **FTS sanitization** — Added input sanitization in `searchFTS()` to prevent query injection
+- **Database** — WAL mode and busy timeout for concurrent access reliability
+
+## [1.1.0] -- 2026-02-23
+
+### Changed
+
+- **Launch Edition** — All features unlocked for free users (unlimited memories, all brain layers, auto-learn, etc.)
+- **URL Migration** — Consolidated all references from `cortex-website-theta.vercel.app` to `cortex-ai-iota.vercel.app`
+
+### Fixed
+
+- **Google OAuth 500 error** — Added `.trim()` to environment variables to strip hidden newline characters
+- **LemonSqueezy webhook** — Updated webhook URL to point to new production domain
+- **Google Cloud Console** — Added new redirect URI for OAuth callback
+
+## [1.0.1] -- 2026-02-19
+
+### Fixed
+
+- **FTS5 indexing** -- Triggers were silently failing to index memories for search
+- **Emoji removal** -- All emojis removed from source code for professional terminal output
+- **SECURITY.md port** -- Documented port corrected from 3100 to 3456
+- **CONTRIBUTING.md** -- Fixed `npm run dev` to `npm run watch`
+- **Config leak** -- Added `bin/` and `.ai/` to `.gitignore` and `.npmignore`
+
+### Added
+
+- **Soak testing** -- 500-operation stability test (1800+ ops/sec, 0 failures)
+- **Binary builds** -- `scripts/build-binaries.js` for standalone executables
+- **CI/CD pipeline** -- `.github/workflows/release.yml` for automated releases
+- **Value verification** -- `scripts/value-test.js` proves cross-session memory recall
+- **Performance stats** -- Added to README (<1ms latency, 1800+ ops/sec)
+- **Direct download** -- README Option 3 for binary downloads
+
+## [1.0.0] -- 2026-02-18
+
+### Added
+
+- **`force_recall` tool** — Mandatory context injection at conversation start
+- **`quick_store` tool** — One-sentence memory storage with auto-classification
+- **`store_memory` / `recall_memory`** — Full memory CRUD with type, tags, reason
+- **`update_memory` / `delete_memory` / `list_memories`** — Memory management
+- **`auto_learn` tool** — Extracts decisions and patterns from AI responses
+- **`verify_code` tool** — Detects hallucinated imports, exports, env variables
+- **`verify_files` tool** — Catches hallucinated file paths
+- **`scan_project` tool** — Scans project structure, stack, and git history
+- **`get_context` tool** — Compressed context for current file
+- **`get_stats` / `health_check`** — Database and server diagnostics
+- **`export_memories` / `import_memories`** — JSON bundle export/import with dedup
+- **Git auto-capture** — `post-commit` hook classifies every commit as a memory
+- **Branch/merge capture** — `post-checkout` and `post-merge` hooks track workflow
+- **Build error capture** — `cortex-run` wrapper captures test/build failures
+- **Optional LLM enhancement** — OpenAI/Anthropic/compatible APIs for smarter classification
+- **Web dashboard** — localhost:3456 with memory viewer, search, export
+- **`cortex-setup` CLI** — Auto-configures Antigravity, Claude, Cursor, Copilot, Windsurf, Zed
+- **CLI routing** — `npx cortex-mcp setup` and `--version` subcommands
+- **Quality gates** — Junk rejection, contradiction detection, deduplication
+- **Rate limiting** — 30 stores, 100 auto-learns, 500 calls per session
+- **Memory decay** — Stale/low-quality memories cleaned automatically
+- **Crash protection** — `uncaughtException` and `unhandledRejection` handlers
+- **Degraded mode** — Server reports errors instead of crashing if DB init fails
+
+### Technical
+
+- SQLite-backed persistence with FTS5 full-text search
+- MiniLM-L6-v2 semantic embeddings via `@xenova/transformers` (background worker)
+- Hybrid retrieval: semantic similarity + keyword FTS + recency scoring
+- Direct SQLite hooks for sub-500ms git hook execution
+- Optional LLM with graceful fallback to keyword classification