Cloudflare Agent Memory: durable recall that skips the context window
The second wave of Agents Week is the infra you actually bind: recall that survives restarts, AI Search freed from AutoRAG config, and Git-scale storage.
Copy markdown
recall(), ingest(), remember() — memory as a Workers binding
Agent Memory is a managed service that extracts facts, events, instructions, and tasks from conversations and returns only what's relevant, so your agent stops re-reading whole histories to stay coherent. It runs five-channel retrieval — full-text, fact-key, raw message, vector, and HyDE — fused with reciprocal-rank fusion, and plugs into the Agents SDK Session API as a model-callable tool. It's in beta.
AI Search: AutoRAG grows up, and it's free for now
The old AutoRAG becomes AI Search — hybrid semantic + keyword retrieval with built-in storage, metadata boosting, and cross-instance search, with no pipeline to wire by hand. It's open beta and free, with at least 30 days' notice before any billing starts, so you can put production-grade RAG behind an agent this weekend at zero cost.
Artifacts: a Git-native filesystem for agent output
Artifacts is a distributed, versioned file system built for agents that speaks Git and scales to millions of repos — the place a code-writing agent can commit without you standing up GitHub plumbing. It's in private beta, with public beta targeted for early this month.
One AI Gateway, 14+ providers; Workflows to 50K concurrency
AI Gateway now fronts 14+ model providers behind a single Workers binding with caching and observability, so switching models becomes a config change instead of a rewrite. Paired with a rearchitected Workflows control plane pushing up to 50,000 concurrent durable runs, it's the plumbing for long-running background agents.
isitagentready.com: a free score for agent-friendliness
A free tool grades how navigable your site is to AI agents, alongside new 'Redirects for AI Training' controls that route verified crawlers to canonical content. If you're building the site an agent has to read, it flags what's breaking before the agent hits it.