Memory MCP Server

# Changelog All notable changes to Memory MCP are documented in this file. The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). ## [0.7.4] - 2026-01-24 ### Added - **Dashboard project filtering** - Memory browser page now includes a project filter dropdown. Select a specific project to view only that project's memories (plus global memories with no project assignment). Useful when working across multiple codebases. ### Changed - **Async mining in hooks** - Both `log-response` (Stop hook) and `pre-compact` (PreCompact hook) now spawn mining as a background process. This eliminates blocking during Claude's operations, making hooks return instantly while mining runs in the background. ## [0.7.3] - 2026-01-24 ### Fixed - **Session project_path bug** - Sessions auto-created via `store_memory()` or `log_output()` now correctly set `project_path` from the current working directory. Previously, sessions created via the internal `_update_session_activity()` upsert had NULL `project_path`, breaking project-filtered session queries. ### Changed - **Removed CI jobs pending secrets** - MCP Registry and Homebrew tap publish jobs removed from `.github/workflows/publish.yml` until repository secrets are configured. PyPI trusted publishing remains active. ## [0.7.2] - 2026-01-23 ### Added - **Beads integration** - New `memory-mcp-cli import-beads` command imports beads issues as memories. Auto-runs `bd export` if available, or accepts piped JSONL input. - **Version bump script** - `scripts/bump-version.sh` updates version in both `pyproject.toml` and `server.json` to keep them in sync. ### Changed - **Dynamic versioning** - `__version__` now reads from package metadata via `importlib.metadata` instead of hardcoded string. Single source of truth in `pyproject.toml`. ### Fixed - **Version sync** - `server.json` now matches `pyproject.toml` version (was stuck at 0.7.0). ## [0.7.1] - 2026-01-23 ### Fixed - **Session linkage for memories** - All memory creation paths now properly link memories to sessions, fixing the dashboard sessions page which showed 0 memories per session. Affected paths: - CLI `seed` command - MCP `seed_from_text` tool - MCP `approve_candidate` tool - Bootstrap from files - Dashboard pattern approval - Mining auto-approve - **Hot cache memory type display** - Fixed `MemoryType.PATTERN` showing instead of `pattern` in the hot cache dashboard page ### Changed - **`get_recent_outputs()` now returns session_id** - The tuple format changed from `(log_id, content, timestamp, project_id)` to `(log_id, content, timestamp, project_id, session_id)` to support session-aware mining ## [0.7.0] - 2026-01-23 ### Changed - **Renamed resources for clarity**: - `memory://working-set` → `memory://hot-cache` (session-aware ~10 items, injected by default) - `memory://hot-cache` → `memory://promoted-memories` (backing store ~20 items, disabled by default) - **Hot cache is now the primary context injection** - The `memory://hot-cache` resource (~10 items) is now the recommended and default context injection. It combines: - Recently recalled memories from the current session - Predicted next memories based on access patterns - Top salience promoted items - **Promoted memories resource disabled by default** - The `memory://promoted-memories` resource returns empty by default. Enable with `MEMORY_MCP_PROMOTED_RESOURCE_ENABLED=true` if you want both resources injected. Promoted memories is the backing store for the hot cache. - **Config setting renames**: - `working_set_enabled` → `hot_cache_enabled` - `working_set_max_items` → `hot_cache_max_items` - `hot_cache_max_items` → `promoted_max_items` - `hot_cache_resource_enabled` → `promoted_resource_enabled` ### Added - **Dashboard hot cache page redesign** - Now shows both hot cache and promoted memories with clear explanations of each: - Hot cache: Session-aware context injected to Claude (recommended) - Promoted memories: Backing store of frequently-used memories - **Config option `promoted_resource_enabled`** - Controls whether `memory://promoted-memories` resource is injected (default: false) - **Backwards compatibility aliases** - Old method names continue to work: - `storage.get_working_set()` → `storage.get_hot_cache()` - `storage.get_hot_memories()` → `storage.get_promoted_memories()` - `storage.get_hot_cache_stats()` → `storage.get_promoted_stats()` ## [0.6.4] - 2026-01-23 ### Added - **Auto-link related memories** - Automatically creates knowledge graph links when storing - New memories are linked to semantically similar existing memories - Configurable: `MEMORY_MCP_AUTO_LINK_ENABLED` (default: true) - Threshold: `MEMORY_MCP_AUTO_LINK_THRESHOLD` (default: 0.75) - Max links: `MEMORY_MCP_AUTO_LINK_MAX` (default: 3) - **Auto-detect contradictions** - Flags potential conflicts automatically - Very similar memories (same topic) are marked as `CONTRADICTS` - Helps identify outdated or conflicting information - Configurable: `MEMORY_MCP_AUTO_DETECT_CONTRADICTIONS` (default: true) - Threshold: `MEMORY_MCP_CONTRADICTION_THRESHOLD` (default: 0.80) - **Dashboard version display** - Shows Memory MCP and MCP versions in footer - **hook-check CLI command** - New command to validate hook dependencies - Checks: uv, jq, database connectivity, hook script, log directory - Supports JSON output for scripting: `memory-mcp-cli --json hook-check` ### Fixed - **Hook script silent failures** - Errors now logged to `~/.memory-mcp/hook.log` - Previously all errors were silently swallowed with `2>/dev/null || true` - Now captures stderr to log file with timestamps for debugging - Refactored CLI invocation into reusable `run_cli()` helper - **Helpfulness/utility tracking** - Fixed broken utility score calculation - Changed `retrieval_auto_mark_used` default to `False` (was `True`) - Previously auto-marked all recalled memories as "used", making utility meaningless - Fixed `retrieved_count` increment in `record_retrieval_event()` - Utility formula `(used_count + 1) / (retrieved_count + 4)` now works correctly - **Hook script project_id bug** - Hook now passes project_id explicitly to CLI - Previously, `cd` to MEMORY_MCP_DIR broke project_id derivation - Now extracts project_path from hook input and passes via `--project-id` - **CLI session_id support** - Hook now passes session_id for provenance tracking - `log-output` and `run-mining` accept `--session-id` option - Hook extracts session_id from input and passes to CLI - **Nested transcript.path parsing** - CLI now handles `{transcript: {path: ...}}` format - Previously only checked `transcript_path` and `transcriptPath` - **server.json version sync** - Version now matches pyproject.toml ## [0.6.2] - 2026-01-23 ### Added - **PreCompact hook** - Automatically save session memories before conversation compaction - New `memory-mcp-cli pre-compact` CLI command - Promotes episodic memories to long-term storage before context is lost - Silent operation - never blocks compaction - **Memory Analyst agent** - Specialized agent for memory health analysis - 5-phase analysis workflow (overview, quality, issues, sessions, recommendations) - Detects contradictions, consolidation opportunities, stale memories - Provides prioritized recommendations (critical/recommended/optional) - Use via Task tool with `subagent_type: "memory-analyst"` ### Fixed - **CI linting** - Formatted 3 files that were failing ruff format check ## [0.6.1] - 2026-01-23 ### Added - **Slash commands for Claude Code** - 14 new `/memory-mcp:*` commands - `/memory-mcp:remember` - Store a new memory - `/memory-mcp:recall` - Search memories semantically - `/memory-mcp:list` - List stored memories - `/memory-mcp:hot-cache` - Manage hot cache (promote/demote/pin/unpin) - `/memory-mcp:stats` - Show memory statistics - `/memory-mcp:bootstrap` - Bootstrap from project docs - `/memory-mcp:link` - Link related memories in knowledge graph - `/memory-mcp:session` - Manage conversation sessions - `/memory-mcp:mining` - Pattern mining from usage - `/memory-mcp:trust` - Manage memory trust scores - `/memory-mcp:consolidate` - Consolidate duplicate memories - `/memory-mcp:forget` - Delete a memory permanently - `/memory-mcp:maintenance` - Run database maintenance - `/memory-mcp:test-all` - Comprehensive interactive testing suite ## [0.6.0] - 2026-01-23 ### Added - **Claude Code plugin support** - Install as a plugin for automatic integration - `.claude-plugin/plugin.json` manifest with MCP server config and hooks - `skills/memory-mcp/SKILL.md` comprehensive usage guide - Install via: `claude plugins add michael-denyer/memory-mcp` - **SessionStart hook** - Auto-bootstrap hot cache from project docs - Runs `memory-mcp-cli bootstrap --quiet` on session start - Automatically seeds from CLAUDE.md, README.md, etc. - **Stop hook with log-response** - Automatic pattern mining from Claude's responses - New `memory-mcp-cli log-response` command for hook integration - Reads transcript, extracts last response, logs for mining - **Bootstrap --quiet flag** - Suppress output for hook usage ## [0.5.17] - 2026-01-23 ### Added - **MLX auto-install on Apple Silicon** - No more manual `[mlx]` extra needed - `mlx-embeddings` now included by default for `darwin + arm64` - 10x faster embeddings work out of the box on M1/M2/M3 Macs ### Changed - **Dashboard timestamp shows time** - Date and HH:MM now displayed in two lines - Previously only showed date; time helps distinguish same-day memories ### Fixed - **vec0 vector insert errors** - Handle orphaned/duplicate vector entries - vec0 extension doesn't support INSERT OR REPLACE - Now DELETE before INSERT in both `store_memory()` and `rebuild_vectors()` - Fixes potential "UNIQUE constraint failed" errors during vector operations ## [0.5.15] - 2026-01-23 ### Added - **Intent-based recall ranking** - Boost memories matching query intent - `infer_query_intent()` detects query purpose (debugging, howto, architecture, etc.) - `compute_intent_boost()` applies category-based ranking boost (up to 15%) - Cheap regex heuristic with zero latency impact - Intents: debugging, howto, architecture, decision, convention, setup, api, todo - **Mining minimum length threshold** - Skip short fragments - `mining_min_pattern_length` config (default: 30 chars) - Prevents storing noise like "Yes" or "OK" as memories - **Dashboard timestamp column** - Memory creation dates visible - "Created" column shows when each memory was stored - Helps distinguish new issues from legacy problems ### Changed - **Command/snippet blocking in mining** - Low-value patterns stored as mined_patterns only - Commands and snippets no longer auto-stored as memories - Still tracked in mined_patterns for reference if needed ## [0.5.14] - 2026-01-23 ### Added - **Injection logging in recall** - Track which memories are retrieved for feedback analysis - `recall()` now logs all returned memories via `log_injections_batch()` - Enables injection pattern analysis and hot cache quality improvement - **Hot cache metric persistence** - Metrics survive server restarts - Promotions, hits, misses, evictions stored in metadata table - Lazy-loaded on first access for performance - **Session summarization tool** - Extract structured knowledge from conversations - `summarize_session(session_id)` groups memories by category - Categories: decisions, insights (lessons/antipatterns/landmines), action items (todos/bugs), context - Returns top 20 items per group sorted by importance - Use before `end_session()` to review what will be promoted ### Changed - **Category gate for auto-promotion** - `command` and `snippet` categories blocked - Low-value categories now rejected with `category_ineligible` reason - `should_promote_category()` helper for promotion decisions - **Knowledge graph links for existing entities** - Entity patterns now tracked even when merged - Fixed bug where semantic dedup prevented relationship creation - Entities linked regardless of `is_new` status ### Fixed - **run_mining KeyError** - MCP tool was reading `new_patterns` instead of `new_memories` - **Staleness thresholds** - Now use `MemoryType` enum keys for type safety - **Recall logging level** - Changed from DEBUG to INFO per changelog claim - **Auto-promotion logging** - Changed from INFO to DEBUG (routine operation) ## [0.5.12] - 2026-01-23 ### Added - **Secret redaction** - Defense-in-depth for output logging - New `redaction.py` module with `may_contain_secrets()` and `redact_secrets()` functions - Secrets redacted BEFORE storage in `log_output()` to prevent persistence - Patterns: API keys (OpenAI, GitHub, AWS), connection strings, bearer tokens, key-value secrets - **Staleness indicators** - Visual hints for memory freshness in hot cache - `_get_staleness_indicator()` returns "never verified" or "stale Xd" based on `last_used_at` - Type-specific thresholds: episodic (3d), conversation (7d), pattern (14d), project (21d), reference (30d) - Integrated into formatted memory output for injection context - **Injection feedback loop** - Hot cache quality improvement from usage patterns - `analyze_injection_patterns()` identifies high-value, low-utility, and promotion candidates - `improve_hot_cache_from_injections()` auto-promotes high-value cold memories - Runs during `run_full_cleanup()` (dry_run=False, actual promotions happen) - **Recall logging enhancements** - Better observability for debugging - Logs query preview (first 50 chars), mode, and elapsed time - INFO level for visibility into recall patterns - **Ranking explanation improvements** - More complete factor breakdown - `build_ranking_factors()` now includes trust_weight, helpfulness_weight - Shows "[keyword boost active]" when hybrid search triggers ### Changed - **Usage-aware trust decay** - Frequently-used memories decay slower - Log-based multiplier: `min(1.5, 1.0 + 0.1 * log(access_count + 1))` - Applied in both recall scoring and hot cache ordering - Rewards memories that prove useful over time - **Hot cache ordering uses decayed trust** - Penalizes stale items - Sort key: session recency → decayed trust → real usage ratio - `real_usage_ratio = used_count / access_count` filters auto-marked noise - Prevents stale high-trust items from dominating injection - **Log level optimization** - Reduced noise, better signal - Eviction: DEBUG → WARNING (cache pressure is notable) - Auto-promotion: INFO → DEBUG (routine operation) - Trust changes: INFO for Δ ≥ 0.1, DEBUG for minor tweaks ## [0.5.11] - 2026-01-23 ### Added - **Promotion rejection tracking** - Observability for hot cache promotion decisions - `record_promotion_rejection()` tracks reasons: category_ineligible, threshold_not_met, low_helpfulness - `get_promotion_rejection_summary()` exposes counts in hot cache stats - Helps tune thresholds and understand promotion patterns - **Exact-match pattern promotion** - Prevent semantic drift in mining - `memory_id` column on mined_patterns (schema v17) links patterns to created memories - `link_pattern_to_memory()` creates the link during pattern storage - Auto-promotion prefers exact match, falls back to semantic search - **Structured logging for scoring functions** - DEBUG logs for all scoring decisions - `infer_category()`, `compute_importance_score()`, `compute_salience_score()` - `get_bayesian_helpfulness()`, `compute_helpfulness_with_decay()` - Parameter breakdown helps debug promotion/ranking decisions ### Changed - **Salience weights rebalanced for usage** - Hot cache now access-heavy - Access: 0.25 → 0.40 (usage is key signal) - Recency: 0.25 → 0.30 (recent patterns more relevant) - Trust: 0.25 → 0.15 (still matters but secondary) - Importance: 0.25 → 0.15 (content value secondary to usage) - **Cold-start threshold lowered** - Faster feedback loop - Bayesian helpfulness gate now triggers at 3 retrievals (was 5) - New memories get benefit of doubt; heavily-retrieved-but-unused blocked ## [0.5.10] - 2026-01-23 ### Added - **Insight extractor** - Extract long-form contextual content from output logs - Pattern matching for architectural decisions, conventions, lessons learned ## [0.5.9] - 2026-01-23 ### Added - **Category-aware helpfulness decay** - Transient categories decay faster - `compute_helpfulness_with_decay()` applies per-category TTL - `todo`, `bug`, `context` decay in 7-14 days - `constraint`, `architecture` persist 60-90 days - Integrated into recall scoring for dynamic ranking - **Auto-pin for high-value memories** - Prevent eviction of critical guardrails - `should_auto_pin()` helper for constraint/antipattern/landmine categories - Requires trust >= 0.8 to auto-pin (validated content only) - `constraint` added to `_HIGH_VALUE_CATEGORIES` for lower promotion threshold - **Low-utility memory penalty** - Auto-weaken frequently-recalled but unused memories - `penalize_low_utility_memories()` runs during maintenance - Trust -= 0.03 when retrieved >= 5 times but used_count = 0 - Helps demote noise from recall results - **Input validation for list_memories** - Reject invalid pagination params - Returns error for offset < 0 or limit < 1 instead of undefined behavior ### Changed - **Trust now affects recall ranking** - `recall_trust_weight` default: 0.0 → 0.1 - Trusted memories rank higher in composite score - Still a small factor vs similarity (0.7 weight) - **LOW_UTILITY trust penalty reduced** - From -0.05 to -0.03 - More gradual decay for low-utility memories ### Documentation - **ML classification guidance** - Added docstring clarifying appropriate uses - Confidence is for category assignment, NOT ranking - Use composite_score for ranking, hot_score for eviction ## [0.5.8] - 2026-01-22 ### Added - **New memory categories** - Better organization of mined patterns - `workflow` - Deployment and operational processes (ssh, curl, deploy scripts) - `snippet` - Code snippets with language markers (never promoted to hot cache) - `command` - Short CLI commands (never promoted to hot cache) - `reference` - External resources, URLs, documentation pointers - `observation` - Factual findings and discoveries - **recategorize CLI command** - Re-run category inference on existing memories - `memory-mcp-cli recategorize --dry-run` to preview changes - `memory-mcp-cli recategorize --all` to recategorize all memories - Useful after category pattern updates - **Auto-mark recalled memories as used** - Automatic helpfulness tracking - New `retrieval_auto_mark_used` setting (default: True) - Memories returned by `recall()` are automatically marked as used - Fixes helpfulness metrics showing 0% (previously required manual `mark_memory_used()` calls) ### Changed - **Low-value categories never promoted** - `command` and `snippet` categories are blocked from hot cache promotion - These are easily discoverable or have low recall value - Keeps hot cache focused on high-value tacit knowledge ### Fixed - **Dashboard sessions page** - Fixed SQL queries using wrong column names - `session_id` → `id`, `created_at` → `started_at` - Sessions page now loads correctly - **Mining button UX** - Added loading spinner and "Running..." text while mining executes ## [0.5.7] - 2026-01-22 ### Added - **Dashboard enhancements** - Full-featured web dashboard at http://127.0.0.1:8765 - **Mining page** (`/mining`) - Review and approve/reject mined patterns - **Injections page** (`/injections`) - Track hot cache/working-set injections over time - **Sessions page** (`/sessions`) - Browse conversation sessions and their memories - **Graph page** (`/graph`) - Knowledge graph visualization with force-directed layout - **Memories over time chart** - Bar chart on overview page showing daily memory counts - **Helpfulness metrics** - Trust score and used/retrieved counts in memory tables - **Category distribution** - Visual breakdown by memory category on overview - **Auto-mining in hook** - Pattern mining now runs automatically after each Claude response - High-confidence patterns stored as memories on first extraction (if content ≥ 20 chars) - Hook processes only text blocks from Claude's response - Hot cache promotion after reaching occurrence threshold ### Changed - **Mining stores memories immediately** - Patterns no longer wait for 3+ occurrences - First extraction creates a memory (if confidence >= threshold) - Hot cache promotion still requires occurrence threshold - Existing patterns migrated to memories on first mining run ## [0.5.6] - 2026-01-22 ### Fixed - **log_output CLI now passes project_id** - Pattern mining was finding 0 outputs because logs were stored without project_id - Root cause: `log-output` CLI wasn't calling `get_current_project_id()` - Mining filters by project_id, so logs without it were invisible - Added regression tests to prevent this from recurring ### Added - **Code map documentation** - Visual architecture guide with bidirectional source links - Mermaid diagrams for system overview, data flows, and schema - Tables mapping components to file:line locations - Quick navigation for key entry points ## [0.5.5] - 2026-01-22 ### Added - **Entity extraction for pattern mining** - Extracts technology and decision entities from Claude outputs - Technology entities: databases, frameworks, languages, tools with context-aware patterns - Decision entities: architecture/design choices with rationale extraction - Confidence scoring based on context quality (rationale, alternatives boost confidence) - **MENTIONS relation type** - New knowledge graph relation for entity linking - Source memories auto-link to extracted technology entities via MENTIONS - Decision entities auto-link to mentioned technologies via DEPENDS_ON - Creates rich semantic connections during mining - **Auto-linking during mining** - `run_mining()` now creates knowledge graph links - Tracks which memories came from each output log via `source_log_id` - Links source content to extracted entities automatically - New helper: `get_memories_by_source_log()` for provenance queries ## [0.5.4] - 2026-01-22 ### Added - **Helpfulness tracking** - Track which memories are actually useful (schema v16) - `retrieved_count` - how often memory appears in recall results - `used_count` - how often memory is marked as helpful - `last_used_at` - when memory was last marked as used - `utility_score` - precomputed Bayesian helpfulness score - **Bayesian helpfulness scoring** - Uses Beta-Binomial posterior `(used + α) / (retrieved + α + β)` - Cold start gives benefit of doubt (0.25) - Low utility detection emerges naturally from retrievals without usage - New helper: `get_bayesian_helpfulness()` in helpers.py - **Helpfulness-weighted recall ranking** - New `recall_helpfulness_weight` config (default 0.05) - `utility_score` now factors into composite recall score - Memories that prove helpful rank higher - **Used-rate promotion gate** - Auto-promotion now requires helpfulness signal - If `retrieved_count >= 5`, requires `used_rate >= 0.25` (25% usage) - Memories below warmup threshold get benefit of doubt - Prevents low-utility memories from being promoted to hot cache - **Hot cache session recency ordering** - Hot memories ordered for optimal injection - Primary: `last_used_at` (most recently helpful first) - Secondary: `trust_score` (reliability) - Tertiary: `last_accessed_at` (general recency) ## [0.5.3] - 2026-01-22 ### Added - **ML-based category classification** - Uses embedding similarity to category prototypes instead of regex - Categories: antipattern, landmine, decision, convention, preference, lesson, constraint, architecture, context, bug, todo - Hybrid approach: ML first, falls back to regex for explicit patterns - Configurable via `ML_CLASSIFICATION_ENABLED` (default: true) and `ML_CLASSIFICATION_THRESHOLD` (default: 0.40) - **Category-aware hot cache thresholds** - High-value categories promoted faster - `antipattern` and `landmine` categories get 0.3 salience threshold (vs 0.5 default) - Temporal-scope-aware demotion: durable categories (2x), stable (1x), transient (0.5x) - **Feedback nudge in hot cache** - Hot cache resource now includes memory IDs and a hint to call `mark_memory_used(memory_id)` when a memory was helpful - **Web dashboard fixes** - Fixed hot cache stats display (`current_count` vs `count`) and pinned status ### Changed - **NER now standard dependency** - `transformers` moved from optional `[ner]` to core dependencies - NER entity extraction enabled by default during pattern mining - Can be disabled via `NER_ENABLED=false` ## [0.5.1] - 2026-01-21 ### Fixed - **Project-scoped mining** - Mining now respects project boundaries - Output logs store `project_id` for filtering (schema v13) - `run_mining` only processes logs from current project - Prevents cross-project pattern leakage and auto-approval - **API endpoint extraction** - Fixed malformed "GET /path /path" patterns - Regex was capturing both full match and path, causing duplication - Now correctly produces "GET /path" format - **Config extraction security** - Hardened against secret leakage - Added `_may_contain_secrets()` filter to config pattern extraction - Tightened regex patterns to capture only safe descriptive values - Prevents sensitive data from being auto-approved to hot cache ## [0.5.0] - 2026-01-21 ### Changed - **Major internal refactoring** - Split large monolithic modules into focused packages - `storage.py` (4,577 lines) → `storage/` package with 16 mixin modules - `server.py` (2,268 lines) → `server/` package with 12 tool modules - No API changes - all imports remain backwards compatible - Each module now follows single responsibility principle - Easier to navigate, test, and maintain ## [0.4.5] - 2026-01-21 ### Added - **Enhanced pattern mining** - Expanded from 5 to 16 pattern types for better extraction coverage - New extractors: decisions, architecture, tech stack, explanations, config, dependencies, API endpoints - NER-based entity extraction (person, organization, location) when `transformers` is installed - Auto-enables NER with `uv tool install hot-memory-mcp[ner]` - no config needed - **Bootstrap context preservation** - Markdown files now preserve section context in chunks - Chunks include source file and section: `[CLAUDE.md > Testing] Use pytest for tests` - Short fact-like items (`Port: 8080`) are preserved instead of being filtered - Non-markdown files get source file prefix: `[README.txt] ...` ### Fixed - **Project-aware deduplication** - Same content in different projects now stays separate - Content hash includes project_id to prevent cross-project merging - Semantic dedup search is now project-scoped - **Recency ranking** - Recall now uses `last_accessed_at` instead of `created_at` - Recently accessed memories rank higher, improving relevance - **Secret detection in mining** - Config extraction no longer captures env var values - Sensitive patterns (passwords, API keys, tokens) are filtered from auto-approval - Only env var names are stored, never values - **NER context** - Named entities now include surrounding context for better recall - Format: `...context... [Entity is a organization]` instead of bare entity - **Mining provenance** - Auto-approved patterns now preserve `source_log_id` for traceability - **Transaction nesting** - Fixed `clear_vectors` to avoid nested transactions ## [0.4.4] - 2026-01-21 ### Fixed - **Hook transcript extraction** - Fixed jq selector to use correct `.message.content` path for Claude Code transcript format. Previously logged raw JSON instead of extracted text. - **Pattern mining auto-approval** - Lowered `mining_auto_approve_confidence` default from 0.8 to 0.5 to match extractor defaults. Patterns meeting occurrence threshold now auto-approve as intended. ## [0.4.3] - 2026-01-21 ### Fixed - Removed unused import in mining module - Use correct settings variable in mining provenance ## [0.4.2] - 2026-01-21 ### Added - **Hybrid search** - Combines semantic similarity with keyword matching for improved recall - FTS5 full-text search table synced with memories via triggers - Boosts results when queries use indirect phrasings (e.g., "FastAPI" matches "framework") - Configurable via `MEMORY_MCP_HYBRID_SEARCH_ENABLED` (default: true) - Adjustable keyword weight and boost threshold settings ### Changed - Database schema version bumped to 12 (auto-migrates existing databases) ## [0.4.1] - 2026-01-21 ### Fixed - Use full server name in mcp-name for registry validation ### Added - Release skill for automated publishing workflow ## [0.4.0] - 2026-01-21 ### Added - **Project awareness** - Memories are automatically tagged with the current git project - Auto-detects project from git remote URL (e.g., `github/owner/repo`) - Recall and hot cache filter to current project + global memories - `memory://project-context` MCP resource shows project-specific context - Configurable via `MEMORY_MCP_PROJECT_AWARENESS_ENABLED` (default: true) - Seamless switching between projects - each sees its own relevant memories - **Episodic memory type** - New `EPISODIC` memory type for session-bound short-term context - `end_session()` tool to consolidate episodic memories at session end - Promotes top memories by salience score to PROJECT or PATTERN type - Configurable retention (7 days default) and trust decay settings - **Working-set resource** - `memory://working-set` MCP resource for session-aware active memory - Combines recently recalled memories, predicted next memories, and top salience hot items - Smaller and more focused than hot-cache for active work context - **Multi-hop recall** - `expand_relations` parameter on recall for associative memory - Traverses knowledge graph relations to find related memories - Score decay for expanded results prevents dilution - Handles SUPERSEDES relation to skip outdated memories - **Consolidation CLI** - `memory-mcp-cli consolidate` command for memory deduplication - Finds clusters of semantically similar memories - Merges near-duplicates, keeping best representative - `--dry-run` flag for preview, `--threshold` for custom similarity - **Unified salience score** - Engram-inspired metric for promotion decisions - Combines importance, trust, access count, and recency - Configurable weights for each component - Used for smarter hot cache promotion alongside access count threshold - **Semantic clustering for display** - RePo-inspired cognitive load reduction - Groups similar memories in hot cache and working set displays - Auto-generates human-readable cluster labels from tags - Configurable threshold (0.70 default), max clusters (5), and min size (2) ### Changed - Hot cache promotion now considers both access count threshold AND salience score - `remember()` tool accepts new `episodic` memory type ## [0.3.0] - 2026-01-19 ### Added - **Research-inspired memory features** - Importance scoring at admission time - Retrieval tracking to learn which memories are actually used - Memory consolidation infrastructure - **Per-result scoring** for recall transparency - Returns similarity, recency, and composite scores in results - Helps understand why memories are ranked the way they are - **Fine-grained trust management** - Contextual reasons for trust adjustments (USED_CORRECTLY, OUTDATED, etc.) - Audit trail for trust changes - Per-memory-type trust decay rates ### Changed - Replaced argparse with click in CLI for better UX - Improved documentation with troubleshooting guide ## [0.2.0] - 2026-01-18 ### Added - **Hot cache truncation** for context efficiency - Configurable max chars per item (default 150) - Prevents context bloat from long memories - **Helper functions module** (`helpers.py`) - Extracted from server.py for cleaner organization - Content summarization, age formatting, confidence helpers - **Database migrations module** (`migrations.py`) - Schema versioning (v1-v10) - Automatic migration on startup - **Response models module** (`responses.py`) - Pydantic models for all MCP tool responses - Better type safety and documentation - **Data models module** (`models.py`) - Enums and dataclasses separated from storage.py - Cleaner imports and organization ### Changed - Refactored server.py to use helper functions - Improved code organization across modules ## [0.1.0] - 2026-01-17 ### Added - **Two-tier memory architecture** - Hot cache with instant recall via MCP resource injection - Cold storage with semantic search via sqlite-vec - **Auto-bootstrap** from project documentation - Detects README.md, CLAUDE.md, CONTRIBUTING.md, etc. - Seeds hot cache when empty - **Pattern mining** from Claude outputs - Extracts imports, commands, project facts - Frequency-based promotion candidates - Human approval workflow - **Knowledge graph** with typed relationships - RELATES_TO, DEPENDS_ON, SUPERSEDES, etc. - Link and unlink memories - **Trust management** - validate_memory() and invalidate_memory() tools - Trust decay over time by memory type - **Session tracking** for provenance - Track which session created each memory - Cross-session pattern detection - **Predictive hot cache warming** - Learns access patterns between memories - Pre-warms cache with predicted next memories - **Apple Silicon optimization** - MLX backend auto-detected on M-series Macs - Falls back to sentence-transformers otherwise - **CLI tools** - `bootstrap` - Seed from project docs - `seed` - Import from file - `log-output` - Log content for mining - `run-mining` - Extract patterns - `db-rebuild-vectors` - Fix dimension mismatches - `status` - Show system health ### Configuration - Environment variables with `MEMORY_MCP_` prefix - Sensible defaults for all settings - Hot cache: 20 items max, 3 access threshold, 14 day demotion - Retrieval: 0.7 confidence threshold, 5 result limit [0.4.2]: https://github.com/michael-denyer/memory-mcp/compare/v0.4.1...v0.4.2 [0.4.1]: https://github.com/michael-denyer/memory-mcp/compare/v0.4.0...v0.4.1 [0.4.0]: https://github.com/michael-denyer/memory-mcp/compare/v0.3.0...v0.4.0 [0.3.0]: https://github.com/michael-denyer/memory-mcp/compare/v0.2.0...v0.3.0 [0.2.0]: https://github.com/michael-denyer/memory-mcp/compare/v0.1.0...v0.2.0 [0.1.0]: https://github.com/michael-denyer/memory-mcp/releases/tag/v0.1.0