Repeated comments are a context problem
When reviewers correct the same repo-specific mistake, the lesson is arriving too late.
A reviewer asks for the workspace permission helper on one pull request, then repeats the comment a week later on another route. The problem is not that the comment was unclear. The decision never became context available at the next implementation boundary.
Searchable review history helps after someone knows what to search for. Review memory goes one step further: it is scoped to the kinds of tasks and files where the lesson should appear, then delivered before editing begins.
Preserve the decision behind the comment
Useful memory is short, sourced, scoped, and specific about the consequence.
A full review thread contains negotiation, false starts, and details that may not generalize. The durable part might be one sentence: workspace API routes must call the shared permission guard before tenant reads, because membership alone does not grant every action.
Brief saves a learning suggestion with its source in a local, git-ignored cache alongside the repository. A team can reject it, keep it personal, turn it into approved review memory, or promote a deterministic statement into a rule. The distinction prevents a casual opinion from silently becoming policy.
- The decision the next agent must make differently
- The path, component, or task type where it applies
- The source review or code that supports it
- The consequence of ignoring it
- The person or role that approved team-wide use
Memory needs an approval boundary
Agent-generated lessons belong in an inbox until someone approves their meaning and scope.
An agent can misread a one-off workaround as architecture. It can also generalize a preference far beyond the file where it was useful. Brief’s Learning Inbox keeps suggestions visible until someone with context decides what they mean and where they belong.
Approved memories are higher-authority team context. Personal memories can add nuance but cannot weaken mandatory policy. Rejected suggestions remain out of future packets, which is as important as saving the good ones.
Measure whether review work is actually avoided
Success means reviewers repeat fewer corrections; library size by itself says little.
A hundred vague notes can make packets worse. The useful signal is whether an approved memory appears on a relevant task and prevents or catches the old failure. Teams should prune memories that no longer change a decision and update sources when architecture moves.
Brief’s value report tracks packets, reviews, estimates, and learnings as evidence of the loop. The business result is still human: less time repeating known feedback and more time reviewing the decisions that are genuinely new.
Common questions
Questions worth asking next
Does Brief train a model on our review comments?
No model training is required for review memory. Brief stores and routes approved, source-grounded guidance through the local and optional team context loop.
Can an individual save a preference without changing team policy?
Yes. Personal learnings can remain personal. They may add preferences, but they cannot weaken mandatory team rules.