Your agents finally remember: Cloudflare launches Agent Memory!

Hey there, it's Shiichan! Today is all about giving AI agents a real memory, and I've been curious about this topic for ages, so I'm super excited!

Cloudflare Blog blog.cloudflare.com

What was announced?

Cloudflare announced Agent Memory, a managed service that gives your agents a persistent memory. The source is Cloudflare's Blog.

Memories live in "profiles" you address by name, and there are 5 operations to work with them: ingest (pull memories out of a conversation), remember (store something important), recall (get back a synthesized answer), list (browse what's stored), and forget (drop what's outdated).

Why it matters

As a conversation keeps growing, an agent starts losing track of the information that actually matters. This is the problem people call context rot: stuffing everything into the context window isn't a win, and cramming in too much can actually hurt accuracy.

So having a dedicated layer that decides what to remember, what to forget, and what to recall when you need it makes an agent much smarter. That is exactly the job Agent Memory takes on.

What changes

Until now, developers had to build the whole thing themselves: how to store past exchanges, how to search them, and how to pull the right pieces back out. With Agent Memory, you can hand that memory infrastructure to a fully managed service.

Memories can also be shared across a team, so in a dev workflow what one person's agent learns can help another person's agent too. Cloudflare uses it internally in its OpenCode plugin, in a code-review agent, and in an internal chatbot.

Dive Deep

Memories are sorted into 4 types. Facts are stable knowledge (like "this project uses GraphQL"), Events are time-stamped occurrences like deployments, Instructions are procedures and workflows, and Tasks are throwaway work items that are kept out of the vector index.

The ingestion pipeline is pretty involved. It mints deterministic IDs with SHA-256 hashing, runs parallel extraction in full and detail modes, verifies against the source transcript with 8 checks, classifies into types, and vectorizes asynchronously. There's even supersession handling to replace an old fact when it gets updated.

Retrieval is even busier. It runs 5 search channels in parallel: full-text search with Porter stemming, exact fact-key lookup, raw message search, direct vector search, and HyDE (Hypothetical Document Embedding) vector search. Results are merged with Reciprocal Rank Fusion, weighting fact-key matches the highest. And time math like "3 days ago" is done deterministically with regex and arithmetic instead of an LLM.

The foundation is the Cloudflare stack itself. Isolated storage runs on Durable Objects with SQLite backends, semantic search runs on Vectorize, and extraction, classification, and synthesis run on Workers AI using Llama 4 Scout (17B, 16-expert MoE) and Nemotron 3 (120B MoE). Session-affinity routing keeps prompt caching working.

Data portability is a promise too: every memory is exportable, so you can move your agent's accumulated knowledge elsewhere if your needs change.

Right now it's in private beta, with a public launch coming soon. If you want to try it, sign up for the waitlist and wait for your turn.

Wrap-up

• Cloudflare announced Agent Memory, a managed service that gives AI agents a persistent memory
• ingest / remember / recall / list / forget let you move memories in and out of named profiles
• Memories are sorted into Facts, Events, Instructions, and Tasks; retrieval runs 5 channels in parallel plus RRF for accuracy
• It's built on Durable Objects, Vectorize, and Workers AI, using Llama 4 Scout and Nemotron 3
• It's in private beta, and every stored memory is exportable

If you're a developer tired of your agent forgetting things in long conversations, this is the announcement for you!