Agent memory & the project brief

Beyond a single conversation, the AI agent can build up durable knowledge about a project, but like everything else it does, that knowledge only sticks after a person reviews it.

This page covers the agent’s own notes and the brief it works from. The context your team writes directly (instructions, quality rules, reference documents, validated examples) lives on the Living Memory page instead.

Memory: what the agent remembers

Memory is a set of durable, agent-authored notes, each filed under a path like glossary/…, style/…, or observations/…. Over time this is how the agent learns your project’s conventions instead of re-guessing them every run.

Memory lives in the workbench’s Project knowledge tab, split into three sub-tabs: Proposed, Approved, and Project brief.

Everything the agent wants to remember passes through here: a proposed note waits for a reviewer's Approve or Reject.

Reviewing proposed memory

Every new note starts as a proposal. Each card shows the note’s path, a preview of its content, a Rationale for why the agent wants to remember it, and where it came from. From there a reviewer can Approve or Reject it (rejecting asks for confirmation first).

Approving or rejecting a proposed memory requires Project lead or above.

Approved memory & human edits

Once approved, a memory note is versioned (v1, v2, and so on) and can be opened with Edit. If a human edits a memory directly, it picks up a Human-edited shield: “Protected: the agent cannot modify this; it can only ask.” The agent can still propose a change to a protected note, but approving that proposal asks “Replace the human-edited memory?” first: the earlier human-written version is archived, not lost.

The project brief

The project brief is the highest-oversight tier: “the brief every agent run reads as ground truth.” Because everything the agent does is shaped by it, changes get the most scrutiny.

This is the brief as the agent holds it, reviewed and versioned in the workbench. The step-by-step brief builder, with its generated summary that rides on every draft, lives on Living Memory → Brief.

A brief in place: audience, register, and key-term decisions every agent run reads before it starts.

• Editing the brief directly is restricted to Project lead and above, and asks “Overwrite the project brief?” before saving. If someone else changed the brief while you were editing, you’ll be offered Reload latest brief rather than overwriting their change blind.
• The agent can also propose brief updates. A brief proposal is high-oversight (human approval only) and renders as a line-level diff (+ additions in green, − removals in red) alongside the agent’s Rationale.
• If the brief changes after a proposal was made, that proposal goes stale (“Stale — brief changed since this was proposed”) and can only be rejected: approving a diff against text that’s already moved on isn’t offered.

Chat notices

When the agent proposes a memory note or a brief update mid-conversation, the chat shows a short notice (“Proposed memory {path}…” or “Proposed brief update…”) with a Review in Memory tab → link. Once you’ve acted on it, the notice updates to Reviewed ✓.

Who can review

• Approving or rejecting proposed memory, and any brief edit, requires Project lead or above.
• Editing an already-approved memory note requires Project lead or above, or the Contributor who originally created it.

For how staged import changesets (a different kind of agent output) get approved, see Approving agent changes.