The Brain · 04.10
Memory in Numbers: Decay, Trust Weighting & Diversification
The scoring mechanics behind recall: freshness decay, trust weighting, reranking, and MMR diversification, explained with real inputs.
The actual numbers, in one place
The rest of this category describes what each mechanism is for. This page is the reference for the specific constants that make recall behave the way it does — useful when you're trying to reason about why a particular memory did or didn't surface.
| Constant | Value | Governs |
|---|---|---|
| Formation minimum score | 0.5 | A statement must score at or above this on the deterministic gate to become a candidate at all. |
| Ask grounding floor | 0.34 | The minimum retrieval score (with lexical-overlap substitution) an atom needs to be allowed to ground an answer. |
| Query overlap minimum | 0.12 | An atom must share at least this much real lexical overlap with the question, independent of its confidence. |
| Semantic dedup threshold | 0.82 cosine | Above this similarity, a new candidate reinforces an existing atom instead of writing a duplicate. |
| MMR λ | 0.72 | How strongly relevance dominates over diversity when selecting which retrieved atoms make the final set. |
| Graduation threshold | 2 retrievals (or 1 reinforcement) | How many times a staged trace must prove useful before it's promoted to durable memory. |
| Procedural TTL | ~60 days | How long an unconsolidated procedural trace survives before expiring unused. |
| Evidence TTL | ~14 days | How long an unconsolidated evidence trace survives — shorter, because bulk evidence is cheaper to re-derive than a hard-won procedural rule. |
Why these are tunable, but not exposed as knobs by default
Every threshold above trades the same two things against each other: how much the Brain remembers versus how confident you can be that what it remembers is actually true and useful. The defaults are set conservative — better to abstain or drop a borderline atom than to pollute recall with something weak. Rerank stays opt-in for the same reason spending stays opt-in everywhere else in Agentis: a cheap heuristic that works is preferred over a model call that isn't obviously worth its cost.
Continue
The background pass that ages, compresses, and reembeds memory over time, plus the health checks that keep recall honest.
The SWIFT verdict engine probes the real world after a run settles instead of trusting a self-report, distinguishing accomplished from merely completed.