Announcing Workspace Memory

Every agent starts the day with amnesia

A coding agent finishes a good piece of work. It learned that migrations are applied by hand after the deploy, and that this repository's tests must run without a database. Then the session ends and all of it is gone. The next agent asks the same question and gets it wrong in the same way.

The usual answer is chat memory: an assistant quietly remembers things about you. That helps one person, not a team. Nobody else can read it, check where a remembered sentence came from, or delete it when it turns out to be wrong. A memory you cannot inspect is a memory you cannot trust with a production repository.

Workspace Memory is our answer. It is available in Aestus today.

Claims, not transcripts

Workspace Memory does not store conversations. It stores claims: one decision, constraint, fact, procedure, or risk at a time, each with a link back to the thing that supports it.

Claims come from work that reached a real state — an accepted outcome, a reworked artifact, a resolved discussion, a handoff comment, a finished coding session — or from a rule a person writes directly. "Someone said this in a chat" carries less weight than "this rule produced an accepted result", and the system treats the two differently.

Because every claim keeps its source, you never have to ask an agent to defend a sentence from memory. You open the ticket, document, or run it came from.

A person keeps the last word

Promotion is automatic when a claim has human-anchored evidence, enough confidence, and a clean safety screen — a queue nobody drains is a feature nobody uses. The stop is just as real: on the Memory screen a person can verify, correct, demote, retract, or purge every claim derived from one bad source in a single audited action. Agent-only proposals stay quarantined and never promote themselves.

The workspace brief has two halves. The team section is written by your team and never rewritten by the system; it is the only part of memory that speaks to agents as instructions. Below it, a generated section is rebuilt from approved claims, and every line footnotes the claims behind it. Its header states its version, its age, and how many claims it covers — so you can see when the brief has fallen behind instead of guessing.

One store, every agent

The point of putting memory in the workspace rather than a chat client is that more than one agent reads it. The brief and the relevant claims now reach board agents, workflows, permitted MCP clients, and governed Claude Code, Codex, and Pi sessions — without crossing the audience boundary a claim was given. A session started through Aestus receives what it is allowed to see, and can propose what it learned, with evidence, at handoff. The channel changes; the claim's audience and citation do not.

That is the result this project was for, and it is now real rather than planned. Before this launch every recorded serving in our own workspace came from Copilot, inside the product. Agent sessions running outside it now read the same claims through the same audited path, and the serving log records which surface received which claim.

What it looks like when it is not working

Memory that only reports success is memory you cannot manage. The Health tab is built to be read when something is wrong.

Four numbers sit across the top. Claims created and servings to an agent are 30-day trends: what the workspace learned, and how often an agent received any of it. Recall abstentions counts the times memory answered "we know nothing about this" rather than returning a weak guess — you want that one visibly non-zero, because a memory that always finds something is feeding agents noise. Dead letters is not a trend: it is the standing backlog of source material that could not be processed, and the only figure printed in red. Underneath, the tab says whether the brief regenerated and whether every claim still carries a current search vector.

We quote no counts of our own here. They move hourly, and a number published on a page is stale the day after it ships. Read yours.

The limits, stated plainly

• Memory is off until you turn it on. Every workspace starts dark; extraction, recall, and every consumer are inert until an admin enables it. There is also a platform kill switch for incident response.
• Candidates still need a person. Automatic promotion needs human-anchored evidence. Anything short of that waits in the queue, and a queue nobody reviews stays a queue.
• Nothing crosses an organization. The organization is the tenant wall. A matching name, identifier, or citation in another organization is invisible.
• We are not claiming an accuracy gain yet. We built the benchmark that compares agent runs with and without memory, and we have not run it. Until it produces a number, we will not tell you agents do better — only that they now have a source.
• Procedural playbooks are not being served. The system induces and stores them, because the benchmark needs the data. Recall withholds them until it shows the arm does not make things worse.

Turning it on

Enable memory for a workspace in its settings, then open Memory in the sidebar. Start with the team section — the rules a new teammate would need on day one. Let accepted results produce the rest.

The test of a memory system is not how much it remembers. It is whether the next agent can act with the team's hard-won context, show where that context came from, and stop using it the moment a person says it is no longer true.