Memory Confidence and Consolidation Lifecycle Automation: A Score the System Can Show Its Work On

A system that flags a pattern without saying how confident it actually is puts the reader in the position of deciding, on faith, whether a two-time coincidence and a five-time habit deserve the same weight. And if that confidence came from a model's read of the situation, asking again tomorrow — same data, same question — isn't guaranteed to produce the same number, which makes the score itself something to double-check rather than something to act on.

Detecting that something recurred is a yes/no question per rule; deciding how much to trust that recurrence is a separate judgment that needs its own explicit math, or it collapses into either flagging everything with the same urgency or leaving the reader to guess how seriously to take it. A confidence number that can't be recomputed from its own inputs — because it came from a model's impression rather than a formula — fails the exact test that makes a trust signal worth having: that two people looking at the same data would arrive at the same number.

Every memory's confidence starts at a base of 22 and adds a fixed amount for each input that strengthens it: 12 points for every occurrence recorded, 11 for every distinct context the pattern has shown up in, 10 for every positive signal attached to it, plus a recency factor that varies by rule. The result is capped to a floor of 28 and a ceiling of 96 — no memory, however thin, scores below 28, and none, however strong, is presented as absolute certainty. That number then sorts into three plain bands: below 58 is low confidence, 58 to 77 is medium, 78 and above is high.

The same numbers also decide which of four lifecycle stages a memory sits in. A brand-new pattern starts "em observação." It becomes "recorrente" once it has at least two occurrences — no confidence minimum required yet, since the point of this stage is just to acknowledge that something has repeated. It becomes "consolidado" once it has at least four occurrences AND at least 72% confidence — both conditions, not either. It only reaches "reutilizável," the stage where other screens are allowed to actively surface it as a working recommendation, once four conditions are true simultaneously: at least five occurrences, at least two distinct contexts, at least 80% confidence, and at least one recorded positive signal.

Nothing here is stored as a finished conclusion. The full list of memories is not read from a saved "memories" table — it is rebuilt from scratch, every time the engine runs, directly from the same live task list, mind maps, decision history and prepared actions the rest of Run already uses. Only the raw events themselves, and how often each memory has actually been reused elsewhere, are persisted; the derived confidence and stage are always a fresh calculation against current data, which means a memory that hit "reutilizável" last week can genuinely drop back down if the underlying signals that earned it change — a completed overdue task, a decision that resolved, a mind map that got real execution behind it.

A recurring-overload pattern for one team first crosses the "recorrente" gate the moment it logs its second overdue-heavy week, with a confidence in the high 40s — no promises yet, just an acknowledged repeat. Three more overdue-heavy weeks later, now with occurrences across more than one context and a couple of resolved incidents counted as positive signals, the same pattern's confidence has climbed past 80 and its stage has advanced to "reutilizável" — the exact point at which it starts actively surfacing as a note inside other screens, rather than sitting quietly in the Memory Engine's own list.

Two patterns with different track records stop looking the same at a glance — the stage and the number both say, in the same fixed language every time, how much weight this specific memory has actually earned. And because the whole list recomputes from live data instead of freezing a past verdict, a memory that stops being true stops being presented as confidently as it once was, without anyone having to go in and manually retire it.

A team wanting the system to explain, in prose, why a specific number is what it is beyond the formula itself will not get an AI-generated justification — the formula is the explanation, and it is the same for every memory. An organization with genuinely little operational history will see most of its detected memories sit in "em observação" or "recorrente" for a while by design — the higher stages are deliberately hard to reach quickly, since they are also the stages other screens are allowed to treat as an active recommendation.

An AI model asked how confident it is in its own read can produce a different-sounding answer to the same question twice, with no arithmetic behind it to check. Centriu Run instead scores every memory with the same fixed formula and the same numeric gates regardless of which organization, which screen, or which day is asking — a score of 80 means exactly the same thing everywhere it appears, and can be recomputed by hand from the occurrences, contexts and positive signals behind it.