What should your AI agents remember?

Memory should be earned, not copied from chat

An agent can sound certain and still be wrong. If every confident sentence becomes memory, one weak guess can return for months. It can enter another task, influence another agent, and slowly become accepted because the system repeats it. A transcript is useful evidence, but it is not a policy manual.

Workspace Memory takes a narrower approach. It turns accepted work and direct human rules into cited claims. A claim is one decision, constraint, fact, procedure, or risk. It links back to the task, document, artifact, or outcome that supports it. That link gives the next person or agent a way to check the source instead of trusting a free-floating summary.

The organization is the tenant wall. Memory from one organization is not served to another. Inside that wall, every claim also has an audience: personal, team, workspace, or organization. The operating cycle has four parts: extract a candidate, promote it when the evidence earns trust, recall only what fits the work, and correct or retire it when the source changes.

1. Extract a claim with its evidence

Extraction starts after useful work reaches a meaningful state. An accepted agent result can teach the workspace what worked. A corrected artifact can teach it what the first attempt missed. A person can also write a rule directly when there is no reason to wait for an agent to infer it.

The extractor does not need the whole conversation. It looks for a statement that can help future work and classifies it:

• A decision records a choice and the path the team will follow.
• A constraint records a boundary the next run must respect.
• A fact records something verifiable about the workspace or product.
• A procedure records how a repeated operation is done.
• A risk records a known failure mode or unresolved concern.

Each candidate keeps its citation. A useful claim might say, “Run database migrations after the deployment reaches Ready.” Its evidence should be the accepted release task or the human-authored runbook that established the rule. “Someone said this in a chat” is weaker than “this rule produced an accepted result.”

2. Promote what deserves to guide future work

New candidates are not all equal. Agent-only proposals stay unreviewed. Automatic promotion requires human-anchored evidence, enough confidence, and a clean safety screen. A person can promote a candidate directly, but the action remains visible in the claim history.

Promotion is the difference between storage and authority. A stored candidate can be inspected. An approved claim can enter an agent's working context. This gate keeps memory useful without pretending the extractor is infallible.

The Memory screen shows the shared brief beside the generated section. The human section is written for the workspace and speaks to agents as instructions. The generated section is rebuilt from approved workspace claims, with a citation on each line. Organization guidance is loaded separately so one workspace does not rewrite the shared organization brief.

Approve claims that are reusable, specific, and supported. Leave one-off details on the original task. Reject a statement that does not help the next agent act or verify.

Choose the smallest useful audience

Audience is part of the claim, not a filter added after recall. Use the narrowest scope that can do the job:

• Personal memory belongs to one person. It is available only when that person is the verified viewer; another user or a service identity cannot inherit it from an assignee field.
• Team memory follows a real team. Current membership is checked when the claim is read, and a non-member receives neither the claim body nor its citations.
• Workspace memory is shared in one workspace. It can guide the people and supported agent surfaces working there, but it does not become visible in a different workspace.
• Organization memory is one shared rule across the organization's workspaces. Only an organization administrator can create or promote it.

These scopes do not weaken the tenant wall. A matching name, identifier, or citation in another organization is still invisible. A useful claim answers two questions: what does it say, and who can receive it?

3. Recall only what fits the work

More context is not always better. A long memory block can hide the rule that matters and use tokens on unrelated history. Recall should select the few claims that match the current task, project, workflow, or question.

When a supported coding session starts through Aestus, it receives the workspace brief and relevant claims allowed for that session. Claude Code, Codex, and Pi sessions can also propose what they learned, with evidence, at handoff. A personal-access-token or verified local session can recall the same permitted memory through MCP. An OAuth-only or service identity does not gain personal or team access merely because it carries a user identifier. The channel changes; the claim's audience and citation do not.

An agent should treat recalled memory as context to verify, not as a command that outranks the current task. It can open the citation, compare the claim with the source, and report when the two disagree. That behavior is why citations matter: they make a compact prompt auditable.

4. Correct the claim, then stop the old one from serving

The source of truth can change. A release process moves. A policy is replaced. A feature that was experimental becomes standard. Memory needs an undo path that does more than edit displayed text.

On a claim, a person can verify, correct, demote, or retract. Correction creates a new current statement and leaves the old state in history. Retraction stops the claim from serving. If one bad source produced several claims, lineage purge retracts everything derived from that source in one audited action.

The detail view keeps the claim, its evidence, usage, confidence, and history together. That makes a correction reviewable before another agent relies on it.

Redaction and consent withdrawal follow the same safety rule. When a source can no longer be served, the source and claims derived from it stop entering new agent context. The audit record remains so an operator can explain what changed and when.

A small operating routine

You do not need a memory-cleanup project before every run. Use a short routine:

1. Put durable rules in the human brief.
2. Let accepted results produce cited candidates.
3. Review candidates that could change future work.
4. Start agents with only relevant approved claims.
5. Correct a claim when its source changes; purge the lineage when the source itself is unsafe.
6. Watch which claims are served, confirmed, refuted, or ignored.

The test for a good memory system is not how much it remembers. It is whether a new agent can act with the same hard-won context as the team, show where that context came from, and stop using it when a person says it is no longer true.