One approved decision layer,
one governed loop.
The right memory goes into the task. What the agent learned comes back — and a human decides what becomes trusted memory for next time.
Nothing agent-generated becomes trusted organizational memory without human approval.
Evidence
Evidence is the raw material a decision gets built from — a meeting transcript, a ticket, a document, an uploaded file, an email thread, a chat excerpt, a PR description. It only enters Decision Memory when someone deliberately adds it or wires up a connector for it.
There's no background crawl of your workspace. If it isn't selected, Decision Memory never sees it — and every piece that is selected keeps its source and timestamp, so a reviewer can always trace a later decision back to where it came from.
Candidate
Once evidence is in, the model reads it and drafts a candidate: what looks like it was decided, why, who was involved, and what it touches. It's a first pass at structure, not a conclusion.
Candidates are labeled as candidates everywhere they show up — in review queues, in search, in agent context — so nothing generated gets mistaken for something the team actually agreed to.
Human approval
A reviewer sees the candidate next to the evidence it came from and decides: approve it as-is, edit the wording or scope, ask for more evidence, or reject it outright.
This gate applies to everything, not just AI-extracted candidates — imported rules from an old .cursorrules or CLAUDE.md file stay candidates too, until a human signs off on them.
Decision memory
An approved candidate becomes a decision record: rationale, scope, owner, status, and a change history that survives past the meeting it came out of.
Records don't disappear when priorities shift — a decision that's later superseded stays in memory with its status updated, so nobody has to reconstruct why something changed.
Agent context
Before an agent starts a task, Decision Memory resolves which approved decisions actually apply — the right scope, the permissions it's allowed to see, and any decisions that conflict with each other.
The agent gets that permitted subset, not the entire decision history. Nothing outside the task's scope, and nothing the agent isn't authorized to read.
Write-back
When the agent finishes, it reports back what it applied, a proposed decision record for anything new, and any open questions it ran into.
That report re-enters the loop as a candidate — never written to memory automatically. It goes through the same human-approval gate as step 03, every time.