Agent Memory

Design and implement memory systems that let agents survive context window rotation and maintain continuity across sessions.

Core Problem

LLM agents have finite context windows. Memory is lost when:

• Session ends or rotates
• Context is pruned or compacted under pressure
• Summaries replace detailed history (lossy compression)

Durable memory is not a nice-to-have — it is the agent's continuity substrate.

Architecture Patterns

Three dominant architectures for persistent agent memory:

1. CMA — Continuous Memory Architecture

Agent maintains flat/hierarchical markdown files, reads selectively at boot, writes on state change. Best for: operational state, ongoing projects, agent identity.