Memory System Overview
How AgentOS memory works end-to-end — classifier-driven dispatch, BM25 + dense + Cohere rerank, Ebbinghaus decay, consolidation — and how each piece earned its place on the 85.6% LongMemEval-S / 70.2% LongMemEval-M result.
Cognitive Memory
Cognitive memory for AI agents: Ebbinghaus decay, HEXACO-modulated encoding, Baddeley working memory, ACT-R spreading activation, 8 neuroscience-grounded mechanisms. 85.6% on LongMemEval-S, 70.2% on LongMemEval-M.
Working Memory
A human-readable .md file that persists across conversations — inspired by Mastra's agent notepad pattern. Complements the Baddeley cognitive working memory with durable, editable state.
Memory Operations
How agents read, write, and serialize their own memory — auto-ingest pipeline, 6 agent-facing memory tools, and import/export across 6 formats (SQLite, JSON, Markdown, Obsidian, ChatGPT, CSV).
Storage & Scaling
Storage adapter quickstart + four-tier scaling path (SQLite → HNSW sidecar → Postgres+pgvector → Qdrant). Covers `@framers/sql-storage-adapter` (better-sqlite3, sql.js, Postgres, Supabase, Capacitor, IndexedDB), cloud backups, and migrations across billions of vectors.