本地记忆管理MCP服务器

memory-mcp

Local memory management MCP server for AI coding assistants (GitHub Copilot CLI, Claude Code).

Indexes MEMORY.md and memory/ files into a SQLite database with hybrid search — BM25 full-text search + vector semantic search — and exposes them as MCP tools that the host CLI calls automatically.

Features

• Hybrid search — combines BM25 keyword matching with vector cosine similarity for best-of-both-worlds retrieval
• Fact + Evidence architecture — memory_write stores compressed facts with linked evidence for traceability
• Semantic dedup — exact string + vector similarity checks prevent duplicate memories
• Memory lifecycle — memory_forget and memory_update with semantic matching (string match → vector fallback)
• CJK support — jieba-wasm pre-segmentation enables Chinese/Japanese/Korean full-text search
• Multi-format — indexes .md, .txt, .json, .jsonl, .yaml, .yml files
• Graceful degradation — vector search requires optional native dependencies; falls back to FTS-only if unavailable
• Session transcript indexing — automatically indexes Copilot CLI (events.jsonl) and Claude Code (~/.claude/projects/) conversation history for cross-session search
• Stat-based fast sync — two-phase sync checks mtime + size first, only reads files that actually changed
• Worker-thread embedding — model loading and inference run in an isolated worker thread, keeping the MCP event loop responsive
• Non-blocking read / blocking write — read tools (search, status) fire-and-forget sync; write tools await fresh sync
• Embedding cache — content-hash keyed cache avoids re-embedding unchanged text across file moves/renames
• Access tracking — frequently accessed chunks get a gentle score boost

Quick Start

# Install globally
npm install -g memory-mcp

# Configure for GitHub Copilot CLI
memory-mcp setup

# Or configure for Claude Code CLI
memory-mcp setup --claude

Next time you launch copilot or claude, the memory server starts automatically.

Vector Search (Optional)

Vector search uses node-llama-cpp and sqlite-vec as optional dependencies. On most platforms (Linux x64/arm64, macOS x64/arm64, Windows x64) these install automatically via prebuilt binaries.

If they fail to install (e.g. missing build tools), the server still works with FTS5-only search — no manual intervention needed.

The default embedding model (embeddinggemma-300M, ~328 MB) is downloaded automatically on first use to ~/.node-llama-cpp/models/. You can swap it via config: memory-mcp config set model /path/to/model.gguf (see Configuration).

Windows: Known Issue with Prebuilt Binaries

The prebuilt @node-llama-cpp/win-x64 binary (compiled with MinGW64) has a bug where getMathCores() returns 0 in certain VM environments (e.g. Hyper-V), causing single-threaded inference (~28x slower than expected). This makes embedding generation too slow to complete within a typical MCP session.

Fix: Build node-llama-cpp from source using MSVC instead of the prebuilt binary:

npx --no node-llama-cpp source download
npx --no node-llama-cpp source build

This compiles with the correct Windows API path and restores full multi-threaded performance. See #1 for details.

Usage

All memory data lives in a shared directory under your home folder:

~/.memory-mcp-workdir/
├── memory-mcp.json              # Configuration file (profiles, settings)
└── default/                     # Default profile workspace
    ├── memory.db                # SQLite database (index + vectors)
    ├── MEMORY.md                # Top-level memory file
    └── memory/
        ├── decisions.md         # Architecture decisions
        ├── preferences.md       # User preferences
        └── evidence/            # Auto-generated evidence files

The host CLI will automatically search these files before answering questions about prior work, decisions, or preferences.

Tools

How it Works

1. The host CLI spawns the MCP server on startup (via stdio)
2. A worker thread is spawned in the background to load the embedding model and handle vector operations — this keeps the main event loop responsive
3. On each tool call, file sync runs in the background (non-blocking for reads like memory_search; blocking for writes like memory_write to ensure dedup sees latest state)
4. Sync uses a stat-based fast-path: checks mtime + size first, only reads and hashes files that actually changed
5. Files are chunked intelligently by format: markdown by headings, JSON by top-level keys/array elements, JSONL by line, YAML by top-level keys or --- document separators
6. Chunks are indexed in SQLite FTS5 (with jieba pre-segmentation for CJK text)
7. Chunks are embedded with embeddinggemma-300M (768-dim vectors) in the worker thread and stored in sqlite-vec
8. Search runs both FTS5 (BM25 ranking) and vector (cosine similarity) in parallel
9. Results are min-max normalized then merged with configurable FTS/vector weighting (default 50/50, see ftsWeight). Vector weight is dynamically scaled by embedding coverage — during index warmup, the system gracefully degrades toward FTS-only
10. Access-count boost gently promotes frequently retrieved chunks: 0.85 × base_score + 0.15 × log₂(1 + access_count) / 10
11. Top-level MEMORY.md / memory.md / MEMORY.txt / memory.txt results receive a score boost (×1.3) for long-term memory priority
12. Falls back to LIKE search if FTS5 is unavailable

Configuration

All settings have sensible defaults and work out of the box. To customize, use the config command:

# Show current config (uses default profile)
memory-mcp config

# Show config for a specific profile
memory-mcp config --profile learning show

# Set a value (on default profile)
memory-mcp config set chunkSize 1024
memory-mcp config set extraDirs /data/obsidian-vault,/data/notes
memory-mcp config set model /path/to/local-model.gguf
memory-mcp config set ftsWeight 0.7
memory-mcp config set minScore 0.05

# Set a value on a specific profile
memory-mcp config --profile learning set chunkSize 256

# Reset to defaults (all profiles and settings)
memory-mcp config reset

# Reset a single profile to defaults (keeps the profile entry)
memory-mcp config --profile learning reset

# Show config file path
memory-mcp config path

Config is stored in ~/.memory-mcp-workdir/memory-mcp.json (cross-platform: $HOME on Linux/macOS, %USERPROFILE% on Windows). Changes take effect on next server restart.

Example Config File

{
  "defaultProfile": "default",
  "profiles": {
    "default": {
      "chunkSize": 512,
      "tokenMax": 4096,
      "ftsWeight": 0.5,
      "minScore": 0.01,
      "maxResults": 10,
      "sessionDays": 30,
      "sessionMax": -1,
      "sessionDirs": [
        { "dir": "/home/user/.copilot/session-state", "kind": "copilot" },
        { "dir": "/home/user/.claude/projects", "kind": "claude" }
      ]
    },
    "learning": {
      "extraDirs": ["/path/to/obsidian-vault"],
      "sessionDirs": [],
      "model": "/path/to/local-model.gguf"
    }
  }
}

Note: Use absolute paths in the config file. ~ is not expanded by Node.js.

Fields omitted from a profile inherit built-in defaults. Each profile workspace is at ~/.memory-mcp-workdir/<profile-name>/ unless overridden.

Config Keys

Profiles

Profiles let you maintain isolated memory spaces for different use cases (e.g. coding vs learning). Each profile has its own workspace, database, and memory files.

# Create profiles
memory-mcp config profile create coding
memory-mcp config profile create learning

# Configure each profile independently
memory-mcp config --profile coding set sessionDirs '[{"dir":"/home/user/.copilot/session-state","kind":"copilot"},{"dir":"/home/user/.claude/projects","kind":"claude"}]'
memory-mcp config --profile learning set extraDirs /path/to/obsidian-vault
memory-mcp config --profile learning set sessionDirs '[]'

# Set which profile is used by default (e.g. by the MCP server)
memory-mcp config profile default coding

# List profiles
memory-mcp config profile list

# Use a specific profile with the CLI
memory-mcp-cli search "query" --profile learning

# Use a specific profile with the MCP server
# In your MCP host config (e.g. .claude.json):
#   "args": ["dist/server.js", "--profile", "coding"]
# Or via environment variable:
#   "env": { "MEMORY_MCP_PROFILE": "coding" }

Each profile workspace lives at ~/.memory-mcp-workdir/<profile-name>/ by default. Deleting a profile removes its config entry but retains the workspace directory on disk — remove it manually if no longer needed.

CLI

A standalone CLI is available for non-MCP integrations (e.g. calling from scripts or other agents):

# Search
memory-mcp-cli search "query" --max-results 10 --min-score 0.1

# Index status
memory-mcp-cli status

# Rebuild the vector index and reclaim disk space
memory-mcp-cli compact

Vector index compaction

sqlite-vec only reuses deleted vector slots in its newest storage block, so edits that replace chunks leave dead blocks behind that every vector query still scans. Indexing only rewrites chunks whose content or position changed, and after each embedding pass the index is rebuilt automatically when fewer than 25% of its slots are live and more than 20,000 are dead.

memory-mcp-cli compact forces that rebuild, then runs VACUUM and prints file size and vector health before and after. If another process is using the database, VACUUM may be skipped; the rebuild still applies. status and memory_status report vecBlocks, vecSlots, vecLive and vecUtilization. When the index holds far more blocks than its live vectors need, search skips vector matching, returns full-text results and logs a hint to run compact.

Search flags

Global config override flags

All config keys can be temporarily overridden via CLI flags (kebab-case):

memory-mcp-cli search "query" --workspace /tmp/alt-memory
memory-mcp-cli status --chunk-size 1024 --session-days 7
memory-mcp-cli search "query" --profile learning --model /path/to/model.gguf

Or run directly: node dist/cli.js search "query"

Output is JSON to stdout, suitable for piping into other tools.

Upgrading

When upgrading from a version that used ~/.copilot/ or ~/.claude/ as the workspace, memory-mcp setup (or the MCP server on first start) will automatically copy your memory files to ~/.memory-mcp-workdir/default/. Original files are preserved — remove them manually when ready.

Config migration: If your config file uses the legacy flat format (no profiles section), it is automatically treated as the default profile. On the first operation that creates profiles (e.g. config profile create), legacy workspace and dbPath fields are migrated into profiles.default so custom paths are preserved.

Breaking changes:

• Most MEMORY_* environment variables have been removed. Configuration is now managed through the config file (~/.memory-mcp-workdir/memory-mcp.json). Use memory-mcp config set <key> <value> to migrate your settings. Two environment variables remain:
  • MEMORY_MCP_PROFILE — select a named profile for the MCP server (alternative to --profile flag)
  • MEMORY_MCP_NO_WORKER=1 — disable the worker-thread embedding engine, fall back to FTS-only search
• The --config CLI flag has been removed. Use --profile to select a named profile instead.

Uninstall

npm uninstall -g memory-mcp

The preuninstall hook automatically removes config and instructions for both Copilot CLI and Claude Code. To remove config manually without uninstalling: memory-mcp uninstall or memory-mcp uninstall --claude

Development

git clone https://github.com/XingxingLi2017/memory-mcp.git
cd memory-mcp
npm install -g .   # Installs dependencies, builds, and registers globally
memory-mcp setup

License

MIT