Memento

Quick Start

Install the plugin, restart your gateway, and Memento starts capturing automatically. Extraction is off by default — enable it explicitly when ready.

Optional: Semantic Search

Download a local embedding model for richer recall:

mkdir -p ~/.node-llama-cpp/models
curl -L -o ~/.node-llama-cpp/models/bge-m3-Q8_0.gguf \
  "https://huggingface.co/gpustack/bge-m3-GGUF/resolve/main/bge-m3-Q8_0.gguf"

Environment Variables

All environment variables are optional — you only need the one matching your chosen LLM provider:

No API key needed for ollama/* models (local inference).

Configuration

Add to your openclaw.json under plugins.entries.memento.config:

{
  "memento": {
    "autoCapture": true,
    "extractionModel": "anthropic/claude-sonnet-4-6",
    "extraction": {
      "autoExtract": true,
      "minTurnsForExtraction": 3
    },
    "recall": {
      "autoRecall": true,
      "maxFacts": 20,
      "crossAgentRecall": true,
      "autoQueryPlanning": false
    }
  }
}

autoExtract: true is an explicit opt-in (default: false). When enabled, conversation segments are sent to the configured extractionModel for LLM-based fact extraction. Omit or set to false to keep everything local.

autoQueryPlanning: true is an explicit opt-in (default: false). When enabled, a fast LLM call runs before each recall search to expand the query with synonyms and identify relevant categories — improving precision at the cost of one extra LLM call per turn.

Data Storage

Memento stores all data locally:

Privacy & Data Flow

Migration (Bootstrap from Existing Memory Files)

Migration is an optional, one-time process to seed Memento from existing agent memory/markdown files. It is user-initiated only — never runs automatically.

What it reads

Migration reads only the files you explicitly list in the config. It does not scan your filesystem, read arbitrary files, or access anything outside the configured paths.

Setup

1. Create ~/.engram/migration-config.json or set MEMENTO_WORKSPACE_MAIN:

{
  "agents": [
    {
      "agentId": "main",
      "workspace": "/path/to/your-workspace",
      "paths": ["MEMORY.md", "memory/*.md"]
    }
  ]
}

2. Always dry-run first to verify exactly which files will be read:

npx tsx src/extraction/migrate.ts --all --dry-run

The dry-run prints every file path it would read — review this before proceeding.

3. Run the actual migration:

npx tsx src/extraction/migrate.ts --all

Security notes

• Migration only reads files matching the glob patterns you configure
• Extracted facts inherit visibility classification (shared/private/secret)
• Secret-classified facts are never sent to cloud LLM providers
• Migration config file is optional — if absent, migration is completely inert
• The migration script has no network access beyond the configured extraction LLM

Architecture

• Capture layer — hooks message:received + message:sent, buffers multi-turn segments
• Extraction layer — async LLM extraction with deduplication, occurrence tracking, temporal state transitions (previous_value), and knowledge graph relations (including causal edges with causal_weight)
• Storage layer — SQLite schema v7 (better-sqlite3) with FTS5 full-text search + optional vector embeddings; knowledge graph (fact_relations with causal_weight), multi-layer clusters, and temporal transition tracking (previous_value)
• Recall layer — optional LLM query planning pre-pass (autoQueryPlanning), multi-factor scoring (recency × frequency × category weight), 1-hop graph traversal with causal edge 1.5× boost, injected via before_prompt_build hook

Requirements

• OpenClaw 2026.2.20+
• Node.js 18+
• An API key for your preferred LLM provider (for extraction — not needed if extraction is disabled or using Ollama)
• Optional: GPU for accelerated embedding search (falls back to CPU gracefully)

Install

# From ClawHub
clawhub install memento

# Or for local development
git clone https://github.com/braibaud/Memento
cd Memento
npm install

Note: better-sqlite3 includes native bindings that compile during npm install. This is expected behavior for SQLite access.