@mem7ai/openclaw-mem7

OpenClaw memory plugin powered by mem7.

Replaces the built-in memory-core with LLM-powered fact extraction, vector + graph dual-path recall, automatic deduplication, and an Ebbinghaus forgetting curve — all running on mem7's Rust core via napi-rs.

Install

openclaw plugins install @mem7ai/openclaw-mem7

Configure

Add the plugin to ~/.openclaw/openclaw.json:

{
  "plugins": {
    "slots": { "memory": "openclaw-mem7" },
    "entries": {
      "openclaw-mem7": {
        "enabled": true,
        "config": {
          "llm": {
            "base_url": "http://localhost:11434/v1",
            "api_key": "ollama",
            "model": "qwen2.5:7b"
          },
          "embedding": {
            "base_url": "http://localhost:11434/v1",
            "api_key": "ollama",
            "model": "mxbai-embed-large",
            "dims": 1024
          },
          "graph": { "provider": "flat" },
          "decay": { "enabled": true }
        }
      }
    }
  }
}

If your OpenClaw config already has an OpenAI provider key (config.models.providers.openai.apiKey), you can omit api_key from the llm and embedding sections — the plugin will resolve them automatically.

Config Reference

| Key | Type | Default | Description | | ----------------- | ------- | -------------------- | --------------------------------------------------------------- | | llm | object | required | LLM for fact extraction and dedup (OpenAI-compatible) | | embedding | object | required | Embedding provider config | | vector | object | { provider: "flat" } | Vector store backend (flat or upstash) | | graph | object | disabled | Graph store backend (flat, neo4j, or kuzu) | | decay | object | { enabled: true } | Forgetting curve parameters | | autoRecall | boolean | true | Inject relevant memories before each agent turn | | autoRecallLimit | integer | 5 | Max memories to inject via auto-recall | | topK | integer | 5 | Default result limit for tool-driven search/list | | searchThreshold | number | unset | Optional minimum score for plugin search/recall | | autoCapture | boolean | true | Extract and store facts after each agent turn | | userId | string | "default" | Base user namespace for long-term memory | | dbPath | string | ~/.openclaw/mem7 | Base directory for SQLite history DB and graph data |

How It Works

Auto-Recall (before_prompt_build / before_agent_start)

Before each agent turn, the plugin:

1. Extracts the latest user message
2. Searches both the current session scope and the broader long-term scope
3. Deduplicates the merged recall set
4. Formats the top memories and graph relations into a context block
5. Injects it as a system prompt prepend

The injected context looks like:

<mem7_context>
## Relevant memories about this user:
- [2026-03-15] Alice loves playing tennis (score: 0.92)
- [2026-03-10] Alice's coach is Sarah (score: 0.87)

## Known relations:
- Alice -[loves_playing]-> tennis
- Alice -[coached_by]-> Sarah
</mem7_context>

Auto-Capture (agent_end)

After each successful turn, the plugin:

1. Extracts the user + assistant messages from the turn
2. Stores them into the current session scope (runId = sessionKey when available)
3. Tags stored memories with metadata.source = "OPENCLAW"
4. New facts are stored; duplicates are merged; stale facts are updated or replaced

This runs as fire-and-forget — errors are logged but never block the response.

Tools

| Tool | Description | | ---------------- | ----------------------------------------------------------------- | | memory_search | Search memories by semantic query, with scope / longTerm | | memory_list | List stored memories for the selected scope | | memory_store | Explicitly store a fact into session or long-term scope | | memory_get | Retrieve a specific memory by ID; path="all" still lists all | | memory_forget | Delete a specific memory by ID, or find delete candidates by query |

Scope Model

The plugin supports three routing modes on search/list/forget tools:

• scope: "session" routes reads/writes to the current session when sessionKey is available.
• scope: "long-term" routes reads/writes to the broader user namespace without a session runId.
• scope: "all" searches or lists both scopes and merges the results.
• longTerm: true is a convenience alias for scope: "long-term".

For memory_store, the default is long-term storage unless you pass longTerm: false or scope: "session".

If the runtime sessionKey matches agent:<agentId>:..., the plugin automatically derives agentId for per-agent isolation. Session-scoped operations use the configured base userId plus the current runId; long-term operations omit runId but preserve the same agentId.

Forgetting Curve

Unlike standalone mem7 where decay is opt-in, the plugin enables decay by default because OpenClaw's long-running sessions accumulate stale facts quickly.

Memories that are frequently recalled decay slower (spaced-repetition effect). Memories that haven't been accessed in weeks are deprioritized but never fully removed.

See the mem7 README for the full mathematical model.

Development

# From the repository root
just openclaw-build
just lint
just typecheck

License

Apache-2.0