Integrations

OpenClaw

Persistent cross-session memory for OpenClaw. Replaces MEMORY.md with a structured knowledge graph that actually scales.

Why Engram + OpenClaw

OpenClaw's built-in MEMORY.md convention is a flat file that grows without bound — it gets stuffed into every session regardless of relevance, becomes expensive and noisy as it grows, and eventually gets truncated. Engram replaces it with a graph-based memory store:

• Retrieves only what's relevant to the current task — not everything you've ever noted
• Scales to thousands of facts without context bloat
• Retires superseded facts automatically — no manual cleanup when decisions change
• Persists across projects with full export/import portability

Token savings compared to a naive MEMORY.md approach: typically 60–80% fewer context tokens per session on a well-established project — exact savings depend on project age and query specificity. Engram's benchmark data is available under benchmarks/ in the repo.

1. Install Engram

pip install engram

Verify the install:

engram --version

2. Add to openclaw.json

In your OpenClaw config file (openclaw.json), add an mcpServers block:

{
  "mcpServers": {
    "engram": {
      "command": "engram",
      "args": ["mcp-serve"],
      "env": {
        "ENGRAM_PROJECT": "my-project"
      }
    }
  }
}

Replace my-project with the name you want for this project's memory store. Engram will create it automatically the first time it's accessed.

Restart OpenClaw. You should see engram_query, engram_extract, engram_stats, and engram_list_projects in the available tools.

3. Update your agent instructions

Replace your existing MEMORY.md instructions in your OpenClaw system prompt or config with:

## Memory (Engram)
At the start of each session, call engram_query with a brief description of
the current task. This retrieves relevant decisions, constraints, and context
from prior sessions — only what's relevant, not everything.

After completing significant work or at the end of a session, call
engram_extract with a summary of what was built, decided, or changed.
Engram will structure and store the facts automatically.

Do not write to MEMORY.md. Use engram_query and engram_extract instead.

4. Migrate existing MEMORY.md (optional)

If you have an existing MEMORY.md, import it into Engram before your next session:

engram init my-project
engram extract my-project MEMORY.md

Engram will parse the file, extract structured facts, and merge them into the project graph. After this, you can leave MEMORY.md in place as a historical artifact or delete it — Engram won't touch it either way.

To verify the import worked:

engram show my-project

How it works in practice

Once configured, your OpenClaw sessions follow this pattern automatically:

1. Session start: OpenClaw calls engram_query with the task description. Engram returns the 10–25 most relevant facts from the project's history.
2. During the session: You work normally. OpenClaw may call engram_query again as the task evolves.
3. Session end: OpenClaw calls engram_extract with a summary. Engram extracts structured facts, detects which supersede prior decisions, and updates the graph.

The graph grows across sessions. Each session starts with a sharp, relevant slice of accumulated project memory rather than a flat growing file.

Hosted API (optional)

By default, Engram runs locally — the memory store is a SQLite file in ~/.engram/. For cross-machine sync or team access, switch to the hosted API:

{
  "mcpServers": {
    "engram": {
      "command": "engram",
      "args": ["mcp-serve"],
      "env": {
        "ENGRAM_PROJECT": "my-project",
        "ENGRAM_API_KEY": "your-api-key",
        "ENGRAM_REMOTE_URL": "https://api.engram.unbidden.ai"
      }
    }
  }
}

Get an API key from the Pro or Team plan. Local mode is free and has no rate limits.

Troubleshooting