Agent Memory System

An evolving architecture for giving long-running agents useful, inspectable memory.

Design principles

The system should be inspectable, selective, and reversible. A human must be able to see why something was remembered and why it returned.

Layers

Working state: the live task and immediate constraints.
Episodes: compact records of decisions, outcomes, and surprises.
Knowledge: stable claims with provenance and confidence.
Policy: rules that govern writing, retrieval, and decay.

Current edge

I’m working on consolidation: deciding when repeated episodes justify a more durable knowledge claim without turning repetition into false confidence.