Most AI memory products answer one question: can we store what you said? That was never the hard part. The hard part is that a memory system sits between four different failures, and a single “confidence score” smears them into one vague warning. A claim can be extracted and still be false. It can be true and still be unwelcome. It can be welcome and still be nobody else’s business.
So Lethe runs four gates, in order, and each fails differently.
Gate one — Found: no quote, no claim
Nothing enters the pipeline as a free-floating impression. Every candidate claim must carry the exact line it came from — your words, verbatim. This is checked in code, not judged by a model: a whitespace-normalised substring match against the source. If the quote isn’t literally there, the claim never existed. In our published benchmark, 95.5% of stored quotes pass this check; the ones that don’t are the first thing we look at when something goes wrong.
Gate two — True: the quote must support the claim
A real quote can still be over-read. “I keep saying I’ll message him this week” supports you postpone a hard conversation; it does not support you dislike him. This gate asks whether the evidence entails the claim — and it’s the one we benchmark hardest: 98.1% of 268 claims across 22 real conversations survived a blind judge. The 1.9% that failed are in the published record, and so is the broken first run of the benchmark itself.
Gate three — Remembered: nothing enters without your yes
A claim that passes both gates is still only a proposal. Lethe shows it to you — accept or deny — and a denied claim doesn’t get stored with a flag on it; it never enters memory at all. This gate is real enough to inconvenience us: the night before recording a demo, a deadline one of us mentioned in chat didn’t pop — because its confirmation was still sitting unaccepted. The gate held against its own builder. Good.
Gate four — Shared: what leaves is a grant, not a copy
Even a remembered fact isn’t automatically anyone else’s to read. Connected AIs get scoped slices — your projects to your coding tool, none of your personal life — through revocable permissions, with every single read written to a log you can open. Raw conversations and the Vault never cross this boundary at all.
Why separate gates matter
Because the remedies are different. A missing quote wants a source. A failed entailment wants the claim corrected or dropped. A denial wants silence. An unscoped request wants a policy answer, not a confidence number. Collapse them and every failure becomes “the AI seemed unsure” — which is how memory products drift into being believed for the wrong reasons.
Found is not true. True is not remembered. Remembered is not shared. A memory you’d let near your life has to keep all four visible.