memento_skill

Overview

This skill provides local persistent memory for OpenClaw agents. It captures conversations, extracts structured facts via a configurable LLM, and auto-injects relevant knowledge before each agent turn. All data is stored locally in SQLite and JSONL to prioritize privacy and offline operation.

How this skill works

Memento listens to message events and buffers multi-turn conversation segments. It optionally sends segments to a configured extraction model to produce structured facts (preferences, decisions, action items, people) and stores facts with provenance and privacy labels. Before each AI turn, Memento recalls relevant facts using FTS5 keyword search and optional semantic embeddings, scoring matches by recency, frequency, and category weight.

When to use it

• Give agents long-term memory of user preferences, decisions, and action items.
• Maintain local-only storage for privacy-sensitive deployments without cloud sync.
• Bootstrap knowledge across agents while keeping private facts siloed.
• Improve agent context with automated recall of relevant facts before responses.
• Migrate existing workspace memory files into a single local store for archives or backups.

Best practices

• Keep autoExtract disabled until you confirm your extraction model and privacy needs.
• Use ollama/<model> for fully air-gapped extraction to ensure no text leaves the machine.
• Tune recall.maxFacts and extraction.minTurnsForExtraction to balance context size and relevance.
• Mark highly sensitive categories as secret so they never participate in extraction or cross-agent recall.
• Run migration dry-runs first to verify which files will be imported.

Example use cases

• A customer support agent that remembers recurring user issues and suggested fixes across sessions.
• Internal team assistants that retain project decisions and action items while keeping HR or payroll notes private.
• An archive/backup tool for all versions of skills and agent conversations stored locally for compliance.
• Offline deployments where embeddings and extraction use local models for semantic recall without cloud access.