OpenClaw¶

Engram has deep integration with OpenClaw — both as an MCP server and via real-time session capture.

MCP Setup¶

Install the engram skill in OpenClaw, then add to OpenClaw's MCP configuration:

{
  "mcpServers": {
    "engram": {
      "command": "engram-mcp",
      "env": { "GEMINI_API_KEY": "your-key" }
    }
  }
}

Session Capture¶

Engram can watch OpenClaw's session files and automatically ingest conversation context into memory. Enable in ~/.engram/config.yaml:

capture:
  openclaw:
    enabled: true
    sessions_dir: ~/.openclaw/workspace/sessions

Start the watcher daemon:

engram watch --daemon
# or as part of the full daemon
engram start

The watcher uses inotify/watchdog to monitor JSONL session files in real time. It captures:

• Text blocks from assistant turns
• Message-sending tool calls
• Skips thinking blocks, trivial messages (< 20 chars), and generic tool calls

How Session Capture Works¶

flowchart LR
    OC["OpenClaw\nSession Files\n(JSONL)"] -->|inotify| Watcher["engram watch"]
    Watcher --> Classifier["Memory Classifier\n(heuristic)"]
    Classifier --> Filter["Entity Gate\n(skip noise)"]
    Filter --> Ingest["EpisodicStore\n+ SemanticGraph"]

Each captured message is classified by type (fact, decision, preference, todo, error, workflow, lesson) using a heuristic regex classifier — no LLM cost at capture time.

OpenClaw Memory Files as Federation Provider¶

Engram auto-discovers OpenClaw's workspace memory files as a federated provider:

# Auto-discovered via file adapter
# ~/.openclaw/workspace/memory/*.md

This means engram recall will also search OpenClaw's memory markdown files alongside engram's own stores.

Per-File Threading¶

The watcher uses per-file threading.Lock to prevent duplicate inotify captures when files are written rapidly.

Verify¶

engram health
engram status

Check that the watcher is running and capturing sessions correctly.