Warp, the terminal company built around AI coding agents, opened a research preview of Agent Memory this week, a system that lets agents keep what they learn instead of forgetting it at the end of every session. The pitch sounds simple: a fact your agent figures out on Monday is available to the same agent on Tuesday, even on a different machine.
That framing undersells what Warp actually shipped. According to Warp’s documentation, memory in this system is not scoped to one person’s tools. It can be organized into stores that belong to a team, then attached to whichever agent everyone on that team already runs, which means a fact learned once can reach every teammate’s sessions without anyone re-teaching it. Convenience across your own machines and convenience across other people’s work are different problems, and Warp’s design treats them differently.
The mechanism starts with how memory gets created. Per Warp, agents write to memory automatically once a conversation ends: the system extracts facts, decisions, and results from the transcript and either merges them into what’s already stored or overwrites it when the new information conflicts. Users can also tell an agent to note something worth keeping mid-conversation. None of it burns tokens or slows down whatever task is actually running, Warp says, because both writing and retrieval happen in the background.
Stores come in three flavors. Personal stores belong to a single user and hold individual preferences and notes. Agent stores are the default: every new agent gets one automatically, though Warp lets a team turn that off at creation. Team stores are the shared kind, meant to hold institutional knowledge like a deployment runbook or a review checklist, and any agent the team already uses can be pointed at one.
Access to a store is not all-or-nothing. Warp lets an administrator attach a store to a given agent as read-only or as full read-write, and it requires a short instruction on every attachment explaining when the agent should consult it. An agent that reviews pull requests can be limited to reading a conventions store it can’t edit, while a deployment agent might get write access to update the same runbook after a rollout. That distinction is the entire difference between an agent that can pollute shared knowledge and one that can only consume it.
Warp also builds in a paper trail. Every memory carries a record of where it came from, and every subsequent edit is logged, so a team can in theory trace a shared fact back to the run that produced it. That matters more than ordinary audit logging, because the correction path here runs through the agents themselves: Warp’s documentation describes memories getting updated and superseded whenever fresh information shows up, even when it conflicts with something stored earlier, and it never mentions a human sign-off before a machine-written memory lands in a team-wide store.
Two capabilities that would harden this system for enterprise use, programmatic API access and self-hosting, are absent from the current preview build. Warp is rolling Agent Memory out to a small set of design partners for now, gated behind a waitlist, and running third-party harnesses like Claude Code and Codex locally against it isn’t supported during this phase; they only get coverage when run as cloud agents.
Teams that get into the preview should treat store creation as an access-control decision, not a convenience setting. Default every team store to read-only until there’s a reason to grant write access, and check whether the audit log is granular enough to delete one bad memory without wiping the shared knowledge everyone else depends on.
Per Warp’s own documentation, published August 18, 2026.
