claw-memory 0.4.0

A lightweight MCP memory server for AI agents. Markdown-first, zero external dependencies.

OpenClaw Memory

A lightweight MCP memory server designed for AI agents. Markdown files as the single source of truth, zero external dependencies.

中文文档

Features

• Markdown-first — All memories stored as human-readable Markdown files, git-friendly
• Zero dependencies — Pure Python + SQLite, no external services required
• Smart writing — Quality gate, auto-routing, conflict detection, reinforcement counting
• Salience-based retrieval — Multi-dimensional scoring: semantic similarity + reinforcement + recency + access frequency
• Token budget aware — Never exceed your context window budget
• Session primer — Cold-start with structured context in ~500 tokens
• Project isolation — Global user memory + per-project working memory
• Privacy protection — Regex-based sensitive information filtering
• V1 zero LLM dependency — Only requires an embedding model (local option available)

Quick Start

1. Install + Initialize (two commands)

# Step 1: Install from PyPI (recommended)
pip install claw-memory[local]          # Local embedding (works offline)
# or pip install claw-memory[openai]   # OpenAI embedding (more accurate)
# or pip install claw-memory[ollama]   # Ollama embedding

# Step 2: Initialize in any project, then restart Cursor
cd /path/to/your/project
claw-memory init

Or install from source (for development):

cd /path/to/claw-memory
pip install -e ".[local]"          # Local embedding
# or pip install -e ".[openai]"   # OpenAI embedding

macOS Note: If you see 'sqlite3.Connection' object has no attribute 'enable_load_extension', your Python was compiled without SQLite extension support. pyenv users can fix this with:

LDFLAGS="-L$(brew --prefix sqlite3)/lib" CPPFLAGS="-I$(brew --prefix sqlite3)/include" \
PYTHON_CONFIGURE_OPTS="--enable-loadable-sqlite-extensions" pyenv install <version> --force

See Troubleshooting for details.

The init command automatically handles all configuration:

• Creates ~/.openclaw_memory/user/ global memory directory with template files
• Creates .openclaw_memory/ project memory directory (journal, agent, etc.)
• Creates .openclaw_memory.toml project config (auto-detects embedding provider)
• Creates .cursor/mcp.json MCP server configuration
• Creates .cursor/rules/memory.mdc agent usage guide
• Creates .openclaw_memory/.gitignore (keeps index files out of git)

Optional flags:

# Specify embedding provider
claw-memory init --provider openai

# Specify project name
claw-memory init --name "my-awesome-project"

# Initialize global memory only (skip project-level files)
claw-memory init --global-only

After init completes, restart Cursor and the agent will automatically use the memory tools.

Other Commands

# Start MCP server (usually called automatically by Cursor)
claw-memory serve

# SSE mode (for web clients)
claw-memory serve --transport sse --port 8765

# One-shot index of existing memory files
claw-memory index

# Open memory viewer in browser
claw-memory web
claw-memory web --port 8767              # Custom port
claw-memory web --no-open                # Don't auto-open browser

Web Memory Viewer

Browse and search all your memories in a clean web UI:

claw-memory web

This opens a local web server at http://127.0.0.1:8767 with:

• File tree navigation — Browse global and project memories organized by category
• Markdown rendering — View memory files with full Markdown rendering and syntax highlighting
• Full-text search — Search across all memory files instantly
• Dark / light theme — Toggle to match your preference
• Metadata display — View frontmatter metadata (type, importance, reinforcement, dates) as badges

Manual Configuration (optional)

If you prefer not to use init, you can manually create .cursor/mcp.json:

{
  "mcpServers": {
    "claw-memory": {
      "command": "python",
      "args": ["-m", "openclaw_memory"],
      "env": {
        "OPENCLAW_EMBEDDING_PROVIDER": "openai",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

MCP Tools

Tool	When to use	Description
memory_primer()	Start of every session	Returns structured context: user identity, project info, preferences, recent activity, active tasks
memory_search(query, scope?, max_tokens?)	When you need to recall specific information	Semantic search with salience scoring and token budget control
memory_log(content, type?)	When you discover information worth remembering	Auto-classifies, deduplicates, detects conflicts, and routes to the right file
memory_session_end(summary)	End of session	Writes structured session summary, updates tasks and primer
memory_update_tasks(tasks_json)	When task status changes	Updates TASKS.md and primer
memory_read(path)	When you need full file content	Reads and returns complete Markdown file

Detailed usage examples: See Cursor Usage Guide for a complete walkthrough of a multi-session scenario showing how the agent uses each tool in practice.

How It Works

Memory Directory Structure

~/.openclaw_memory/              # Global (cross-project)
├── config.toml                  # Global configuration
├── user/
│   ├── preferences.md           # Your preferences
│   ├── instructions.md          # Your rules for the agent
│   └── entities.md              # People, tools, projects
└── index.db                     # Global vector index

<project>/.openclaw_memory/      # Per-project
├── .openclaw_memory.toml        # Project configuration
├── PRIMER.md                    # Auto-maintained session primer
├── TASKS.md                     # Active task tracking
├── journal/YYYY-MM-DD.md        # Structured daily session logs
├── agent/
│   ├── patterns.md              # Reusable solution patterns
│   └── decisions.md             # Architecture decisions (ADRs)
└── index.db                     # Project vector index

Smart Writing Pipeline

Input --> Quality Gate --> Privacy Filter --> Smart Router --> Reinforcement/Conflict Check --> Write

1. Quality Gate: Rejects noise (too short, filler phrases, pure code, speculation)
2. Privacy Filter: Blocks API keys, passwords, internal IPs (configurable regex)
3. Smart Router: Auto-classifies content to the right file by keyword patterns
4. Reinforcement: If highly similar memory exists (>0.92), increments reinforcement count instead of duplicating
5. Conflict Detection: If similar memory exists (0.85-0.92) with new info, replaces the old entry

Salience-Based Retrieval

salience = 0.50 * semantic_similarity
         + 0.20 * reinforcement_score
         + 0.20 * recency_decay
         + 0.10 * access_frequency

Memories that are frequently mentioned (high reinforcement), recently updated, and often recalled naturally rank higher — no manual importance tuning needed.

Token Budget

# Returns as many relevant memories as fit within 1500 tokens
results = memory_search("webhook handling", max_tokens=1500)

Configuration

Project config (.openclaw_memory.toml)

[project]
name = "my-project"
description = "E-commerce platform"

[embedding]
provider = "openai"              # openai | ollama | local
model = "text-embedding-3-small" # optional, uses provider default

[privacy]
enabled = true
patterns = [
    'sk-[a-zA-Z0-9]{20,}',
    'ghp_[a-zA-Z0-9]{36}',
    'password\s*[:=]\s*\S+',
]

[search]
default_max_tokens = 1500
recency_half_life_days = 30

Environment Variables

Variable	Description	Default
OPENCLAW_EMBEDDING_PROVIDER	Embedding provider	local
OPENCLAW_EMBEDDING_MODEL	Model name	Provider default
OPENCLAW_MEMORY_ROOT	Override memory root path	Auto-detect
OPENAI_API_KEY	OpenAI API key	—

Embedding Providers

Provider	Dimension	Dependency	Use case
OpenAI text-embedding-3-small	1536	API key	Best accuracy
Ollama nomic-embed-text	768	Local Ollama	Offline / privacy
sentence-transformers all-MiniLM-L6-v2	384	Pure local	Zero dependency

Design Decisions

This project was designed by analyzing four existing memory systems:

• memsearch: Markdown as source of truth, content-hash dedup, hybrid search
• OpenViking: Directory-based organization, L0/L1/L2 progressive loading
• memU: Reinforcement counting for importance, salience scoring formula
• claude-mem: Structured session summaries, project isolation, privacy tags

Key differences from all four:

• Zero LLM dependency in V1 (only embedding model needed)
• Zero external service dependency (pure Python + SQLite)
• Smart writing pipeline replaces LLM-based extraction with rule-based routing
• Reinforcement + rules hybrid for importance (data-driven + heuristic)
• Token budget aware retrieval (none of the four do this)

License

Apache-2.0. See LICENSE for details.