Memory and Storage

Agents get better over time because they keep two kinds of state: memory (what they’ve learned) and storage (the files they work with). Both are inspectable on the agent’s page.

This is what separates the tenth run of a job from the first. Without persistent state, every run starts from a blank slate and you re-brief the agent on the same quirks — the fiscal calendar, the report format you actually want, which alerts are noise. With it, corrections stick, working files carry forward, and the agent you have in month three is measurably more useful than the one you created. The two kinds of state divide cleanly: memory holds facts and lessons, the things an agent should know; storage holds files, the things an agent is working on.

Memory

The Memory card holds “knowledge this agent saves across conversations.” Each entry shows when it was created and updated and how often it has been used, and expands to show its full value. Every entry keeps revision history — open View history to see how a memory evolved and restore an earlier version if an update went wrong.

In practice, entries read like the notes a careful new hire keeps. After a few weeks of running a Monday pipeline report, an agent’s memory might contain:

• “The team’s fiscal year starts in February; quarterly summaries should break at February, May, August, and November — not calendar quarters.”
• “Dana wants the pipeline report as a single table with week-over-week deltas, no executive-summary paragraph.”
• “Deals whose champion has left almost never close; flag them separately rather than counting them in the weighted total.”

Each of those was earned once — from a correction, an answered question, a pattern the agent noticed — and applies to every run after.

You don’t write memories by hand; the agent stores them as it learns from conversations. Correct it in conversation (“actually, our fiscal year starts in February”) and the correction becomes memory.

Correct in conversation, not in the output

When an agent gets something wrong, the instinct is to fix the deliverable and move on. Say it to the agent instead — “we report in fiscal quarters, starting February” — and the fix becomes memory, so you never make the same correction twice. This is also the remedy for a wrong memory: state the correction plainly, or restore an earlier version from the entry’s history.

Dreams

Once a day, while idle, the agent consolidates what it has learned and logs it in the Dreams tab. Each dream is an audit of the consolidation — what was merged, promoted, updated, or forgotten — so even the agent’s housekeeping is transparent.

Consolidation is what keeps a busy agent’s memory from silting up, and the Dreams tab means it never happens invisibly. If an entry you cared about changed overnight, the dream that changed it says so, and the entry’s revision history lets you put it back.

Storage

The Storage card shows the agent’s durable working storage: “the agent’s working storage is snapshotted each session; everything outside of it is ephemeral.” You can see the home directory’s size and last snapshot time, browse the snapshot’s files with read-only previews, and review snapshot history. Deleting a snapshot rolls back to the previous one; deleting the storage entirely gives the agent a fresh start.

Storage is where an agent’s work-in-progress survives between sessions — the draft it was iterating on, a dataset it assembled over several runs, the scratch notes an assignment builds up week over week. The snapshot browser makes that state as inspectable as memory: when you wonder what an agent has actually been accumulating, open the card and look.

Snapshot history doubles as an undo. If a run leaves the files in a bad state — a draft overwritten, a dataset mangled — deleting the offending snapshot rolls back to the one before it. And because memory and storage are independent kinds of state, tending one never requires touching the other.

Keeping an agent’s state healthy

• Teach through conversation — corrections said to the agent become memory; corrections made silently to its output do not.
• Skim the Memory card periodically — especially for a shared agent your team relies on, the entries are a fast read on what it believes.
• Read the dream after an intense week — a burst of sessions means a burst of learning, and the consolidation log shows what stuck.
• Prefer restore to re-teaching — when a memory update or a snapshot went wrong, rolling back is faster than explaining everything again.

Whether an agent may form memories at all is one of the self-improvement guardrails, so an agent’s capacity to learn is itself something you control.