> ## Documentation Index > Fetch the complete documentation index at: https://docs.plyra.dev/llms.txt > Use this file to discover all available pages before exploring further. # Overview > Persistent structured memory for AI agents AI agents forget everything between sessions. Context windows overflow. Users explain themselves again and again. plyra-memory fixes that. It gives your agent a persistent, three-layer cognitive memory — working, episodic, and semantic — with a hybrid retrieval engine that fuses results across all three, ranked by similarity, recency, and importance.
Plyra Memory Architecture
## Three memory layers Current session scratchpad. Fast writes, in-memory. Flushed to episodic at session end. Event timeline with full conversation and action history. Vector search across all past sessions. Structured knowledge graph — subject, predicate, object. Facts that survive restarts and are never duplicated. ## Local-first. Server-optional. By default, plyra-memory runs entirely on your machine — SQLite for structured data, ChromaDB for vectors. No external services, no API keys. When you need multi-agent memory sharing or multi-tenant isolation, point `PLYRA_SERVER_URL` at a [plyra-memory-server](/memory-server/overview) instance. Zero code changes — the library detects the env var and routes automatically. Remember your first fact in 60 seconds. Deep dive into working, episodic, and semantic storage.