Finding something is easy. Finding everything is the job.
Most memory systems measure whether they retrieved something relevant. Regulated work needs to know they retrieved everything relevant — because the one record that slips through is the one the auditor finds. That gap has a name: completeness. It is the metric FerriteMem is built for.
The two questions
they look identical; they are notEvery memory system answers one of two questions. One asks whether anything relevant came back. The other asks whether everything relevant came back. On the same retrieval, over the same records, these produce completely different verdicts — and the difference between them is the entire regulated market.
“Find me something”
the chatbot question — did anything relevant come back?
One relevant result is a win. Perfect for search and assistants: you asked a question and got a useful answer. If a second relevant item existed and never surfaced, no harm done.
- one relevant piece found
- three others missed, unnoticed
“Find me everything”
the audit question — did everything relevant come back?
Every relevant result must surface. Miss one and the answer is wrong, however many you found — because the missing piece is exactly the one a regulator, a court or a clinician will ask about.
- three of four found
- the fourth decides the outcome
The field optimises and publishes the left panel — it is the right metric for search and chat. Regulated work only pays for the right one.
Where FerriteMem stands
the strict metric, on a public benchmark, at full scale94.8% of the time, everything relevant comes back.
On a public long-memory benchmark across the full 500 questions, FerriteMem retrieves everything relevant 94.8% of the time, and something relevant 99.4% of the time.
We lead with the hard number — the one that fails when anything is missed — because it is the one that matters when the answer has to hold up to review.
The full results table| Metric | Result | What it asks |
|---|---|---|
| recall_all@10 | 0.948 | was everything relevant retrieved |
| recall_any@10 | 0.994 | was anything relevant retrieved |
Completeness is only half the problem
finding every record is worthless if you trust the wrong oneRetrieve all six records perfectly — and if the system believes the one that was later corrected, the answer is still wrong. When two stored facts contradict each other, something has to decide which one wins. Most systems hand that decision to a language model. FerriteMem decides it deterministically, with no model in the loop.
A model picks
a language model reads both and chooses
Two stored records disagree: a cap of 10M recorded in January, a cap of 5M recorded in March. The model reads both and answers “probably 5M”.
- not reproducible — may differ next run
- not explainable — no stated reason
- can be tricked by a crafted record
FerriteMem decides by rule
a fixed rule resolves it, deterministically
The same two records. A stated rule selects the current one and reports the alternative alongside it, so the reasoning is visible rather than inferred.
- reproducible — the same answer, always
- explainable — the rule is the reason
- cannot be tricked — there is no model to fool
The superseded January record is never deleted. It stays, marked superseded, for the audit trail. And a malicious “ignore the safety rule” record cannot quietly win, because a person approves every change to the lasting record and no model does the choosing.
Two halves, one design
both follow from the same choiceOn the way out
Completeness — find every relevant record, not merely a plausible one. Measured on the strict metric rather than the forgiving one.
On the way in
Deterministic resolution — decide conflicts by rule rather than by model, so the record the system trusts is always explainable.
Both come from the same architectural choice: no language model in the core. Together they are what lets an answer survive being checked.
See what a miss costs, industry by industry