Agents That Remember: Introducing Agent Memory

Cloudflare believes that proper memory management matters more than a larger context window for agents to produce better outcomes. Even with a window exceeding 1 million tokens, the context rot problem remains, and both extremes—stuffing everything in or discarding everything—have their limits.

So Cloudflare has released Agent Memory as a private beta. It's a managed service that extracts information from conversations, stores it in a separate memory profile, and retrieves it again when needed. Profiles are addressed by name, and the core operations are divided into ingest, remember, recall, list, and forget.

The architecture is retrieval-based, and the key point is that it doesn't let the agent design its own storage strategy. Cloudflare explains that this constraint is better not only for cost and performance, but also for reasoning problems in production environments such as temporal logic, supersession, and instruction following. It also emphasizes that memory is exportable, so users won't be locked into the vendor.

Here's how it works:

• During compaction, the harness sends the conversation, triggering bulk ingestion, which extracts facts, events, instructions, and tasks from the messages.
• Each message receives a content-addressed ID hashed from the session ID, role, and content, making duplicate ingestion idempotent.
• The extractor runs a full pass over roughly 10K characters at a time, along with a detail pass that fills in specific values for longer conversations, running in parallel.
• The verifier validates facts through 8 checks, including entity, location, temporal, organizational, and relational context.
• Facts and instructions are grouped by topic key, and when a new entry comes in, it supersedes the existing memory rather than deleting it.

Retrieval is also composed of multiple channels. Porter stemming-based full-text search, exact fact-key lookup, raw message search, direct vector search, and HyDE vector search all run in parallel, then are combined via Reciprocal Rank Fusion (RRF). Notably, fact-key matches receive the highest weight, while raw messages remain as a safety net to catch clues that extraction may have missed.

The use cases are broad. It supports coding agents like Claude Code, custom harnesses, long-running background agents, and even shared memory at the team level. As a result, the experience of a single person or a single agent is no longer confined to a one-off conversation, but accumulates as a reusable organizational asset.