agent-memory-architectureslisted

# Persistent Memory Architectures for Agents An agent that forgets everything between sessions can't learn, personalize, or maintain long projects. This skill covers memory as a system architecture: layered stores, retrieval pipelines, consolidation jobs, and forgetting policies — protocol- and vendor-neutral. ## Overview Think of agent memory like a data platform with three layers. The hot layer is working context: fast, exact, bounded. The warm layer is searchable episodic and semantic memory: past interactions, facts, preferences, indexed for similarity and keyword retrieval. The cold layer is consolidated knowledge: distilled playbooks, long-term project state, archived with pointers back to sources. Between layers run pipelines — consolidation (compressing hot into warm/cold) and retrieval (surfacing warm/cold into hot at decision time). Forgetting is a first-class pipeline too. ## When to use - Designing memory for an agent that works with a user or project over weeks or months. - Choosing between memory approaches: flat vector store vs. layered architecture vs. structured records. - Debugging "the agent forgot" or "the agent remembers wrong" problems. - Adding personalization or long-horizon project tracking to an existing agent. ## Core concepts - **Layered stores**: working (session), episodic (what happened), semantic (what's true), procedural (how to do things). Different lifetimes, different retrieval, different update rules. - **Memory records**: