Warp documented Agent Memory, a research-preview persistent memory system that lives on Warp rather than inside a single agent. It is shared across every supported harness — the built-in Warp Agent, Claude Code, Codex and others as they are added — and agents read from and write to it as they run, so durable facts, decisions and outcomes from one conversation are available to the next regardless of which harness, machine or teammate triggers the work. Creation and retrieval run asynchronously in the background, and access is enabled per team for design partners from a waitlist.
Putting memory outside the harness is the interesting choice. It makes the memory a shared team asset rather than a per-tool convenience, and it also relocates the governance question: whatever is written becomes visible to the next harness that reads it, including one a different person started.
That is exactly the property tested by the agent lock-in test being a full session export, and it is the write-loop argument recorded as a pattern in company memory needing write loops. The agent memory hub tracks where that state is allowed to live.
This is research-preview documentation gated to design partners: it describes behaviour without published retention, scoping or redaction rules for what agents write. The condition to watch is whether memory writes become reviewable before they are shared, because a cross-harness memory that accepts whatever one agent concluded propagates its mistakes just as efficiently as its facts.