Memento

Observational memory for OpenClaw.

Most agent memory systems get the abstraction wrong — they either replay full transcripts (too noisy) or constantly reinject changing summaries (churn cache, fragment attention). Memento takes a different path.

It watches sessions as they happen, extracts genuinely notable observations — decisions, constraints, preferences — and maintains a compact, prioritized log that persists across sessions. Memory is split into two scopes: shared (follows the agent everywhere) and session (scoped to the current conversation). As observations accumulate, Memento periodically consolidates them, dropping stale context and preserving what matters. At the right boundaries, it feeds just enough back into context to maintain continuity without dragging history through every turn.

No vectors, no databases. Just a Markdown file that LLMs already understand.

Install

Use OpenClaw's plugin installer (recommended):

openclaw plugins install @makeseabed/memento

If you're running from a local OpenClaw checkout:

pnpm openclaw plugins install @makeseabed/memento

In most setups, that is enough to register and enable the plugin.

Configuration

All config is optional. If omitted, Memento uses built-in defaults.

{
  "model": "openai/gpt-5-mini",
  "observer": {
    "maxSessions": 10,
    "maxLinesPerTranscript": 300,
    "existingObservationsContext": 40
  },
  "watcher": {
    "turnThreshold": 10
  },
  "reflector": {
    "triggerWordThreshold": 8000,
    "backupRetentionCount": 10
  },
  "memoryFlush": {
    "flushLookbackHours": 2,
    "recoverLookbackHours": 4,
    "skipDedup": true
  },
  "logging": false
}

Config surface

• model: shared model for observer and reflector unless overridden below
• observer.maxSessions: recent sessions to scan
• observer.maxLinesPerTranscript: transcript lines to read per session
• observer.existingObservationsContext: recent memory bullets used for dedup context
• observer.model: observer-specific model override
• watcher.turnThreshold: run observation after this many meaningful assistant replies written to the session transcript
• reflector.triggerWordThreshold: consolidate when memory gets too long
• reflector.backupRetentionCount: reflector backups to keep
• reflector.model: reflector-specific model override
• memoryFlush.flushLookbackHours: lookback window for pre-compaction capture
• memoryFlush.recoverLookbackHours: lookback window for reset recovery
• memoryFlush.skipDedup: skip dedup during flush and recovery flows
• logging: write memento/memento.log when true, off by default

Environment overrides

Environment variable overrides are still supported in the current codepath and take precedence over config. The primary env key is now MEMENTO_LOGGING, with legacy MEMENTO_LOG_FILE_ENABLED still accepted for compatibility.

Examples:

• MEMENTO_MODEL
• MEMENTO_OBSERVER_MODEL
• MEMENTO_REFLECTOR_MODEL
• MEMENTO_WATCHER_TURN_THRESHOLD
• MEMENTO_REFLECTOR_TRIGGER_WORD_THRESHOLD
• MEMENTO_LOGGING

The source also supports overrides for the other numeric and boolean settings in src/config.ts.

Files

Memento organizes memory under each agent's workspace:

Shared memory lands in memento/shared/ — durable observations that follow the agent everywhere.

Session memory lands in memento/sessions/<stableChatKey>/ — local observations scoped to a specific conversation. The chat key is derived from OpenClaw's real session key as a stable dashed name, for example agent-main-discord-channel-1480872431068516454.

Each store contains observations.md, a backups/ directory, and a last-observed-at timestamp. Agent-level runtime files live at memento/memento.log and memento/.observer-state.json.

Contributing

Memento is open source and contributions are welcome.

• Open an issue before starting significant work so we can align on direction.
• Keep PRs focused — one fix or feature per PR.
• Tests are required. Run pnpm test before submitting. For bugs, include the relevant section of memento/memento.log and your OpenClaw version.

Inspiration

Memento stands on the shoulders of some excellent work in the agent memory space.

Mastra AI's observational memory planted the conceptual seed, demonstrating that agents need more than context windows; they need a way to observe, distill, and remember what matters.

Gavdalf's article clarified why the observation layer matters and helped shape the framing behind Memento.

Gavdalf's total-recall then provided practical scaffolding. Its prompt engineering and code patterns directly influenced Memento's implementation.

This project is our own take on the problem, but the conceptual and practical debt is real, and we're grateful for it.

License

MIT