tiered-memory
por bowen31337
Coding Agents & IDEsResumen
EvoClaw Tiered Memory Architecture v2.2.0 - LLM-powered three-tier memory system with automatic daily note.
Descripción
Qué hace esta skill
Three-tier memory system (hot/warm/cold) for persistent agent memory across conversations. Facts are distilled, scored with recency decay, and organized in a hierarchical tree index. Critical state syncs to Turso after each conversation for cloud-first recovery and multi-device continuity.
LLM-powered tree search achieves 98%+ retrieval accuracy versus 70-80% for vector embeddings, and context size stays constant at 8-15KB regardless of how long the agent has been running.
Cuándo usarla
- Retrieving an architectural decision made three weeks ago without scanning full history
- Automatically ingesting daily markdown notes into the tiered memory system
- Finding all stored URLs associated with a specific tool or project
- Keeping agent context under 15KB after years of accumulated conversations
- Restoring full agent memory to a new device in under two minutes
Flujo de trabajo de ejemplo
User asks: What did we decide about the BSC integration approach?
- 1
Agent sends the query to the tree search module
- 2
LLM reasons over the tree index and identifies relevant categories such as projects/evoclaw/bsc and active_context/events
- 3
Warm memory entries from those categories are fetched and ranked by score
- 4
Cold storage is queried for lower-scored archived facts in the same paths
- 5
Top results are injected into the agent context before the response is generated
Fact returned: 'Decided to use raw JSON-RPC for BSC to avoid go-ethereum dependency' with category path, importance score, and any preserved URLs or commands
Requisitos
- Turso account with a database URL and auth token (TURSO_URL and TURSO_TOKEN env vars) for cold storage and critical sync
- An LLM API endpoint or local model server for LLM-powered distillation and tree search (rule-based mode works without one)
Instalación
clawhub install bowen31337/tiered-memoryEjecuta el comando de la CLI o pega el prompt en Claude Code, Cursor, Codex o el agente de programación que uses.