Offloop
Sign in

AI agents that remember the work

Work continues from what the team already settled, instead of restarting at an empty prompt.

Get startedHow memory works

Continue the work without reconstructing it

The useful unit of memory is not every token an agent has seen. It is the part of the work a team would recognize and an agent is allowed to use.

Goal

What the team is trying to make true.

Decisions

What was approved, changed, or ruled out.

Evidence

The files and results that support the work.

History

What already happened, and what remains.

The next run starts with the last decision attached

A team should be able to resume from the project state, not paste yesterday’s conversation back into a new window.

See how teammates continue the same work

Without agent memory

Jun 11, 9:02 AM

New chat for launch week. Recapping: fall campaign, we compared three cuts last week, paid social won, budget stays at $40k.

Jun 11, 9:04 AM

Understood. Which file is the current brief — and was the landing page approved, or is that still open?

With Agent memory

Paid social wins — the numbers aren't close. Lock it in.

Jun 4, 9:36 AM

Agreed. @Launch Manager lock it and cap the budget at $40k.

Jun 4, 9:38 AM

Locked. The brief carries the decision, the cap, and the one open step.

Worked for 4 minutes
Jun 11, 9:40 AM
Jun 11, 9:40 AM

@Launch Manager draft Friday's announcement post.

Jun 11, 9:42 AM
Reading last week's brief and the locked decision…

Draft ready — it follows the paid-social direction locked in launch-brief.md, inside the $40k cap.

Worked for 2 minutes

Three steps decide whether the next task starts informed or blank

Each one leaves something a person can open, so what an agent carries forward stays checkable rather than implied.

01

Capture

Every run writes its outcome back into the channel that asked for it.

02

Retrieve

The next task pulls back only the records it is authorized to open.

03

Continue

Work resumes with the prior outcome and the next step visible to the team.

Agent memory is not unlimited recall.

What memory covers

  • Scoped — memory belongs to the work and the workspace that created it.
  • Relevant — retrieval brings what supports the current task, not every prior message.

What memory answers to

  • Permission-aware — losing access to the source must not create new authority.
  • Inspectable — the record stays visible to the people responsible for it.

Questions about agent memory

What memory means, what persists, and what it does not imply.

Let the next task begin where the last one ended.

Get started