agent-memory-systems · davila7 — Claude Code skill: agent memory and retrieval

What agent-memory-systems does

This skill treats agent memory as a system rather than a bucket. It walks through memory types: short-term (the context window), long-term (vector stores), working, episodic, semantic and procedural memory. Memory formation, retrieval and decay each get their own pass.

The author's core claim is that a memory failure reads as an intelligence failure. When an agent forgets something or contradicts itself, retrieval is usually the culprit, not storage. So the skill covers patterns for picking a vector store, strategies for chunking documents, and anti-patterns such as keeping everything forever or chunking without testing retrieval.

A sharp-edges table rates each problem by severity: contextual chunking, testing chunk sizes, filtering by metadata before the search, temporal scoring, conflict detection on write, token budgets per memory type, and recording the embedding model in metadata. Related skills it points to are autonomous-agents, multi-agent-orchestration, llm-architect and agent-tool-builder. Build agent memory or RAG without retrieval tests and you will hit these same walls.