total-recall
por gavdalf
DevOps & CloudResumen
The only memory skill that watches.
Descripción
Qué hace esta skill
Autonomous memory layer for AI agents that runs on a cron schedule, compressing conversation transcripts into prioritized notes without manual saves. Five redundancy layers cover cron-based observation, automatic consolidation, session recovery, a reactive file watcher on Linux, and a pre-compaction hook. An optional nightly Dream Cycle archives stale memories and adds semantic retrieval hooks.
Unlike memory skills that require explicit save commands or vector databases, it observes and compresses sessions automatically on a schedule with no user action required.
Cuándo usarla
- Picking up context from a previous coding session without re-explaining the project
- Tracking evolving decisions across multi-day agent conversations
- Recovering observations from sessions the cron job missed
- Running a low-cost background memory system on a local or cloud agent
- Keeping agent startup context lean with nightly consolidation of a growing observation log
Flujo de trabajo de ejemplo
User starts a new agent session after working on a project the previous day
- 1
Session Recovery runs at /new and checks whether the previous session was captured by the Observer
- 2
Detects a hash mismatch and triggers Observer in recovery mode with a 4-hour lookback
- 3
Observer reads recent JSONL transcripts, deduplicates via MD5 hash, and sends messages to the LLM
- 4
LLM compresses the conversation into prioritized observations appended to observations.md
- 5
Agent loads observations.md at startup and resumes with full prior context
Agent has accurate cross-session memory of decisions, preferences, and in-progress work from the previous day
Requisitos
- OPENROUTER_API_KEY environment variable (OpenRouter account required)
Instalación
clawhub install gavdalf/total-recallEjecuta el comando de la CLI o pega el prompt en Claude Code, Cursor, Codex o el agente de programación que uses.