04 · Memory

Memory.

Agents write down what the next agent should know. In the default review mode, nothing they write is trusted until you publish it.

The problem.

The next worker believes the last note. A wiki and a folder of notes both answer with something. One confident, wrong note becomes ground truth for every agent after it.

What it is.

Agents write down decisions, lessons, and conventions. Each note starts as a candidate and stays out of agent search until it is published.

A judge scores candidates. Projects set to auto or trusted publish some without you. A lesson pulled from a closed item always waits for you, in every mode, because an item's description is written before the work and can be contradicted by what shipped.

Example · project Core Platform
agent:auto-extractLesson from CORE-12Candidatejudge: not judged
agent:grillDecision from CORE-P3Candidate
user:alexAPI errors carry a hint fieldPublished

candidate → you · publish / reject → published

What it isn't.

Not a judge deciding what is true. A model can score candidates. In the default mode, publishing is yours.

Memory in the product Why a folder of notes cannot say it does not know

What's new

Every entry is work that merged. The full history is on GitHub.