Agents that remember: introducing Agent Memory

Cloudflare has announced the private beta of Agent Memory, a managed service that gives AI agents persistent, retrievable memory across sessions. It addresses the context rot problem by extracting facts, events, instructions, and tasks from agent conversations during compaction, storing them in isolated profiles, and surfacing relevant information on demand via a multi-channel retrieval pipeline. The architecture uses Durable Objects for isolated SQLite storage, Vectorize for vector search, and Workers AI for LLM inference. Retrieval combines full-text search, exact fact-key lookup, direct vector search, HyDE vector search, and raw message search, fused via Reciprocal Rank Fusion. Memory profiles can be shared across agents, team members, and tools, enabling institutional knowledge accumulation. Cloudflare uses it internally for coding agents, agentic code review, and chat bots. Data is fully exportable to avoid vendor lock-in.