memory-mimir

OpenClaw plugin that connects to Mimir — a unified long-term memory engine. Gives your OpenClaw agent persistent, searchable memory across sessions.

How It Works

User message → auto-recall (search Mimir) → inject <memories> → Agent runs → auto-capture (ingest new messages)

Auto-recall: Before each agent turn, extracts keywords from the user message (supports English + CJK), searches Mimir with RRF hybrid retrieval (BM25 + vector fusion), and injects the most relevant memories into the LLM context.

Auto-capture: After each agent turn, incrementally captures new conversation messages and sends them to Mimir for extraction (episodes, entities, relations, events).

Features

• Agentic search with query classification and strategy routing
• CJK (Chinese/Japanese/Korean) keyword extraction
• Time range extraction from natural language ("last week", "昨天")
• Score-ordered compact memory formatting (2000 char budget)
• Incremental capture — no duplicate ingestion across agent turns
• File attachment capture — automatically extracts and stores images/documents from conversations

Setup

Prerequisites

A running Mimir server.

Install

npm install
npm run build

Disable OpenClaw built-in memory (if configured)

If you have OpenClaw's memory-core extension active (i.e. you configured an embedding provider), disable it to avoid conflicts:

{
  "plugins": {
    "memory-core": { "enabled": false }
  }
}

Note: Most users don't have memory-core configured, so this step is usually unnecessary.

Configure

In your OpenClaw plugin config:

{
  "mimirUrl": "http://localhost:8766",
  "userId": "your-user-id",
  "groupId": "your-group-id",
  "autoRecall": true,
  "autoCapture": true,
  "maxRecallItems": 25,
  "maxRecallTokens": 2500
}

| Option | Type | Default | Description | |--------|------|---------|-------------| | mimirUrl | string | — | Mimir server URL | | userId | string | — | User ID for memory isolation | | groupId | string | — | Group ID for memory scoping | | autoRecall | boolean | true | Inject memories before each turn | | autoCapture | boolean | true | Capture conversations after each turn | | maxRecallItems | number | 25 | Max memory items per recall | | maxRecallTokens | number | 2500 | Max token budget for memory context | | displayName | string | — | Display name in captured messages (defaults to userId) |

Architecture

memory-mimir (this plugin)          Mimir-Go (engine)
┌─────────────────────┐            ┌──────────────────────┐
│  extractKeywords()  │            │  Agentic Search      │
│  extractTimeRange() │──search──→ │  ├─ Query Analyzer   │
│  formatResults()    │←─results── │  ├─ BM25 + Vector    │
│                     │            │  └─ Graph Traverse   │
│  auto-capture       │──ingest──→ │  Pipeline            │
│  (incremental)      │            │  ├─ Narrative Extract │
│                     │            │  └─ Graph Extract    │
└─────────────────────┘            └──────────────────────┘

Development

# Run tests
npm test

# Build
npm run build

License

MIT