Governed memory
FuseOne memory stores structured assertions, not free-form prose that the model is expected to obey later.
An assertion has a kind, subject, signature and claim. It also carries evidence, labels, confirmation state, retention and erasure state. The model can search memory, but it does not choose the company, area or agent namespace being searched.
When a person teaches from a run, they write only a short subject, claim and
reason, then explicitly choose agent or shared reach. The platform records new
teaching as kind fact and derives the signature from the trimmed subject. The
derived identity is visible before saving and cannot be edited; the agent,
labels, citation, counters and expiry come from the ledger and policy.
The subject is the memory’s stable human key, so different facts need different
specific subjects.
Reading memory can taint the run
Section titled “Reading memory can taint the run”Memory does not make outside data trusted. If an assertion came from Slack, GitHub, Loki or another external source, its labels travel with the memory read. The fold applies those labels to the run, and the Gate evaluates the next write with them.
That is enforcement, not a prompt instruction.
Agent-proposed memory
Section titled “Agent-proposed memory”Agents can suggest memory only when the published version opts in. In review mode, the suggestion enters the review queue without an extra runtime approval; a person accepts or dismisses it before it becomes active. In auto-confirm mode, the same assertion must appear across the configured number of distinct runs before it becomes active, so tainted suggestions can still require Gate approval before they count.
The agent proposes only the subject and claim. The platform derives the same
fact/subject identity used by human teaching, so active memory and pending
proposals share one duplicate-detection space.
The model does not send confidence. Counts are derived by the platform.
Active memory is corrected through the governed record path, not edited in place. A correction rewrites the claim, keeps the same signature and evidence, and records a new event with the person’s reason. If the signature is wrong, disable the assertion and record a new one.
Search and budget
Section titled “Search and budget”Memory search is indexed for broad substring matching, but responses are still bounded. If matching assertions do not fit the response budget, the tool says how many were omitted and why. Labels are unioned across all matching assertions before that response cut, so an omitted risky assertion cannot make the run look cleaner than it is.