Memory Pattern — Short and Long-Term Agent State

Why it matters

Memoryless agents forget everything. Well-designed memory unlocks 'agents that know me' experiences.

The architecture

Short-term: current turn context. Message history plus tool results.

Long-term: semantic store. RAG retrieval per turn.

How it works end to end

Context management: as short-term grows, summarize or truncate. Prevents token overflow.

Memory curation: agent decides what to promote to long-term. Simple heuristic or LLM-driven.

Retrieval: embed current query; find similar long-term facts; inject.