UniloomDocs
Use Uniloom

Tidying memory

Memory grows with every session, so Uniloom keeps it small and right without a model of its own. Your agent proposes tidies, AI memory care can check memories each night, and every change can be undone.

A tidy#

When a project's memory needs it, a session offers to tidy it, at most once a week. If you agree, the agent reads the project's memories and proposes:

  • a merge: several memories become one;
  • a retirement: a memory that is wrong or outdated;
  • a lesson: several memories become one rule with a when (Lessons).

The change applies at once. Tidied this week, at the top of To review, says what tidies changed in the last 7 days, with one Undo. A memory you pinned or protected is never touched. A change the save rules refuse waits under To review for you instead.

AI memory care#

AI memory care checks your memories each night with your own agent. Turn it on at the end of Built-in in Automations. It applies to this workspace only.

  • At 03:00, when the computer is awake and Uniloom is open, it reads up to 50 memories of one project or group, with the code each cites. A missed night is skipped.
  • It reads each memory twice, separately, and retires only what both readings call wrong. When the readings disagree, the memory waits under To review.
  • Before you turn it on, it shows what the first month will cost in tokens. It stops for the month at the Monthly cap you set.
  • A memory it retired can be brought back from Retired for 30 days.
  • While it is on, an agent's new memory must say in its reason why it changes what a future agent does.

For your own schedule instead, add the Tidy memories starter in Automations.

Lessons from what you decline#

Uniloom keeps your no to an AI proposal: an automation's held call answered No, a step sent back, a widget card declined, a Home tool call skipped. A no becomes a proposal under To review when you wrote a reason in Why? (optional), or when you declined the same action twice in 30 days in the same place. Accepting it saves a lesson. At most 5 such proposals wait per project.

When a lesson becomes a skill#

A lesson used in at least 3 sessions is a way of working. Uniloom proposes it as a skill in Toolkit › Suggestions, at most 3 a day. Make it a skill writes the skill into the project's folder, where every agent loads it, and the lesson leaves the memory index. A dismissed proposal is not asked again. A lesson with no project folder on this computer is never proposed.

Does memory help?#

The Does memory help? card on Activity (add it while arranging the board) compares sessions given memory with the one session in four held out. It counts the turns where you corrected the agent ("no, …", "wrong", "I told you"), per session, in each group. It shows a result once each group has 50 sessions. You can turn the holdout off there.

Two more cards read the memory layer: Do sessions save memory? and Is memory used?.

Last updated