Agent Memory Architecture

A memory architecture for AI agents that separates four distinct memory types and handles each correctly, rather than collapsing everything into a single MEMORY.md.

The four types: working memory (current session context), episodic memory (records of past conversations), semantic memory (extracted facts), and procedural memory (skills and workflows). Each has a different file format, update frequency, retrieval pattern, and size limit.

The architecture: MEMORY.md holds semantic facts with a pruning trigger. SESSION-LOG.md holds episodic records with a rolling window. CONTEXT.md holds working state that resets each session. Each file has an update protocol and a clear owner.

What is included: the full architecture guide, file templates for all four memory types, the AGENTS.md memory integration block, a pruning script that archives old session logs, and the size limits and update triggers for each file.

Best for operators whose agents are forgetting things, loading irrelevant context, or behaving inconsistently across sessions.