tiered-memory

Organize AI agent memory across hot, warm, and cold storage tiers by structuring identity, active context, recent facts, and long-term archives into agent-scoped JSON files. This structured memory system replaces flat note files to help agents retain important facts and discard noise reliably across sessions.

Tiered-memory does not strictly require a database connection to store agent context. It operates using Python standard library tooling and agent-scoped JSON files, while offering optional Turso access via libsql_client and optional LLM endpoints for cloud-first sync and advanced processing.

Extract structured facts from raw LLM conversations using LLM-powered distillation and tree search. This mechanism turns raw daily note ingestion into compact facts and retrieves relevant memory categories by reasoning, preserving URLs, commands, and file paths during metadata extraction.

The best way to validate agent memory and track health metrics is using a system that warns when notes are incomplete, applies score-based pruning, and monitors health metrics. This preserves critical state and maintains agent consistency over long-lived workflows like EvoClaw and OpenClaw.

Yes, you can sync agent memory state to the cloud for distributed workflows. The system supports cloud-first sync via optional Turso access using the libsql_client dependency, keeping critical state safe and accessible across distributed agent environments.

You should avoid a tiered memory approach if your agent operates in stateless, single-session environments where long-term context retention is unnecessary. The system is designed for long-lived work requiring daily note ingestion, consolidation, and context recovery, adding overhead for simple, ephemeral tasks.