Work memory (Lab)

Work memory is an optional Lab feature that keeps the durable things your work produces — decisions, learnings, outcomes — so the next session, teammate, or AI agent doesn't have to rediscover them. Your AI tool proposes a memory as it works; a human confirms it; from then on, agents read it back as settled fact.

The timeline already records everything you did. Work memory is the selective layer on top: the rare thing worth remembering about it.

⚠️ Pulse is in early beta, and Work memory is a Lab feature. It's off by default, and it may change. Switching it off hides everything again — nothing already captured is deleted. Today's flow is agent-suggests, human-confirms; expect more ways to capture a memory over time, potentially including a conversational way to add one directly, not only through an agent's tool call.

How it works

An agent proposes. While working, your AI tool notices something durable and files it as a suggestion.
A human reviews. You open Work memory from the account menu and confirm, edit, or dismiss it.
Agents read it back. Confirmed memories are searchable by every agent with access, so the team builds on past decisions instead of re-litigating them.

🛑 Agents can never confirm a memory — by design. There is no "confirm" tool in MCP or REST. Confirmation is what turns an AI's inference into shared truth, so it stays a deliberate action in the app. The one exception is when you explicitly tell your agent to remember something ("remember that we don't use Supabase Auth here") — that saves as confirmed, because a human already decided it.

Memory types

Type	What it captures
Decision	An explicit choice, or a resolved comparison — "X rather than Y".
Learning	A discovered constraint, surprising behavior, or a failed approach with a reusable lesson.
Outcome	The result of something — a migration finished, an experiment's verdict.
Milestone	A checkpoint reached — a launch, a deploy.
Blocker	Something that stopped progress, and why.

Each memory has a title (the decision or lesson itself), a one-or-two-sentence summary that still makes sense a year from now, optional detail, the date it happened, an optional project, and — for agent-captured ones — a confidence score and links back to the time entries it came from.

Routine activity — fixing a typo, running tests, an ordinary commit — is deliberately not a memory. Agents are told to stay selective.

Turn it on (Admin)

Work memory is off until an Admin enables it.

Open Settings → Lab.
Set Work memory to Admins only or Everyone.
- Admins only is the way to trial it: you get the screen and the tools, nobody else sees a change.
- Everyone gives every member the screen and the tools.
That adds a Work memory screen to the account menu and the memory tools to your MCP server. Reconnect or restart your AI tool if it doesn't list the new tools right away.

Once it's on, two more settings appear:

Agent capture — whose AI agents may capture memories through MCP and REST (Off / Admins only / Everyone, default Everyone). This is separate from who can use the screen, because capture and review are different audiences. A common setup: Work memory = Admins only and Agent capture = Everyone, so every engineer's agent files suggestions while only admins triage them. Agent capture can never be wider in effect than the main setting.