Reasera Memory
A memory your AI can be held to.
Give agents and assistants a store of what your organization knows — with the discipline to answer only from evidence, to say “no record” when there is none, and to unlearn on command. Every answer comes with a receipt.
What every answer comes with
A certainty level, a receipt, or a reason.
Certainty level
Your application sets the bar for each use: what to show, what to caveat, what to escalate.
Receipt
Receipts make answers checkable by people and auditable after the fact.
Reason for a gap
Certified absence
“No record” is an answer, not a shrug.
A vector store cannot tell you something is not there — a low similarity score is a guess. Reasera keeps a complete index of everything in memory, so absence is a checkable certificate. Your agent can say “we have nothing on this” and mean it.A certificate says nothing in memory covers this — not that nothing exists anywhere. Because the index is complete rather than a similarity search, a record filed under that subject cannot be missed. If the record is there under a different name, the result is a gap with a reason, never an invented answer, and it lands on the gap list so a person can connect the two.
90%
declined when the subject’s document was in memory but did not contain the answer (159 of 176). Standard RAG cannot decline: it answered all 176, 22% of them wrongly.
Measured on our evaluation panels · How we measure
Retraction that survives rewording
Retract a claim once. In 13,090 trials, none came back.
Retraction is about the claim, not the document. Delete a document from a vector store and a paraphrased copy will serve the same wrong claim again tomorrow. Retract it in Reasera and reworded copies are refused as well, including copies that arrive later, while unrelated facts keep serving. When the answer has simply changed, supersede it with the new one and the old answer stops serving.
100%
of withdrawn claims stayed withdrawn after the same claim came back in different words (13,090 of 13,090 trials). Standard RAG, which deleted the document instead: 47%.
Measured on our evaluation panels · How we measure
Couldn’t you just delete it?
We did, on the RAG side — the claim came back anyway. Deleting removes a document; a wrong claim lives in many documents, and new copies keep arriving in tickets, transcripts, and re-crawls. Retraction is a standing instruction: this claim is withdrawn, in whatever words it appears, now and later. And unlike deletion, it keeps the record of what was claimed, when it was withdrawn, and by whom.
Erasure that is real
When content must not exist, it doesn’t.
Retraction says the claim is wrong: the content stays for the record, and memory refuses to serve it. Redaction says the content must not exist: it leaves the ledger and every cached copy, and the store is rebuilt without it. Both actions require a person with the right role — API keys cannot perform them.
Redaction removes
- ✓ the source text
- ✓ its span in the store
- ✓ every cached copy
What remains
- • the event: who, when, which record
- • a one-way fingerprint for the audit trail
Gaps are a to-do list
Every decline points at the missing source.
Each gap is recorded with its subject and reason, surfaced on your dashboard, and pushed to a webhook the first time it is seen. Your content team gets a ranked list of exactly what memory is being asked and cannot yet certify — the shortest list of what to add next.
Memory gaps · this week
| Q3 price list | no source | 14× |
| Northgate SLA credits | conflicting sources | 6× |
| EU data residency | insufficient evidence | 3× |
Measured
The usual alternatives, on the same questions.
Same small language model, same documents, both sides. Standard RAG — retrieval plus free generation — against Reasera Memory. Retrieval found the right document almost every time on both sides; the difference is what the model is allowed to say next.
When it answers, how often is it right?
Questions the model could not answer from its own knowledge, with the answer present in the documents. Retrieval found the right document 98% of the time for both sides. Reasera gives 53% of them a certain answer, flags another 32% as lower-confidence, and declines the rest with a reason. (9,792 questions)
Does a correction stick?
Withdrawn claims that stayed withdrawn after the same claim came back in different words and the question was asked again. Reasera retracted the claim; RAG deleted the document instead. Unrelated facts blocked: 12 of 39,270 checks. (6,545 claims, two rewordings each)
Does it admit when it has nothing?
Questions where the subject’s document was in memory but did not contain the answer. Reasera declined 159 of 176 at the certain level and answered 17 wrongly. RAG cannot decline: it answered all 176 from the model’s own recollection — unverifiable, and wrong 22% of the time. (176 questions)
Do you need the big model?
How often it is right when it answers, same questions. The comparison model is sixty times larger, and with RAG it is still bound by what free generation produces. (9,792 questions)
Panels drawn from all 21,919 CounterFact records, one per subject, with Wikipedia summaries as the documents: 9,792 facts the base model gets wrong with the answer present; 176 whose document lacks the answer; 6,545 retractions with two rewordings each. Both sides use the same open 124M-parameter model; the last card compares against an 8-billion-parameter model. Standard RAG = dense embedding retrieval, top document from a 10,650-document pool, free greedy decoding; the same word-level scoring for every arm. Excluding the 1,095 questions whose answer appears in the subject’s own name: 80% vs 22%. Answers flagged lower-confidence are right 53% of the time; your application decides whether to show, caveat, or verify them. On facts both sides already know, the two tie (90% vs 89%). A fixed sample of these panels gates every release.
Results are from Reasera’s fixed evaluation panels, described above, and are re-measured before each release. Performance on your data depends on your sources, configuration, and questions, and is not guaranteed.
How we measure. Questions come from CounterFact, a public dataset of facts, with each subject’s Wikipedia summary as the document. A question is assigned to a panel by how the base model behaves on its own, and retraction trials re-add a served fact in the dataset’s own rewordings. Both sides use the same model and the same scoring rule; Reasera’s headline figures count only answers given at the certain level. We publish our misses: on the declining panel they were near-synonyms and fragments of the subject’s own name.
Integration
An afternoon, not a migration.
Memory is an HTTP API and an MCP server. Add sources, ask questions, check for absence, and read the gap list from any programming language. Agents and copilots connect through MCP — the standard way agents plug in tools — and get the same receipts.
POST /ingest— add a document, transcript, or recordPOST /recall— ask; get a tier, an answer, a receipt or a reasonGET /absence— certify that memory holds nothing on a subjectGET /gaps— what memory is being asked and cannot certifyretract · supersede · redact— curation, people only
API keys carry memory-specific scopes that generic read/write keys do not imply. Curation actions are never available to keys. Webhook: memory.gap_detected.
Questions are phrased as statements to complete. Memory finds the source from the question and names it in the response; pass a subject to ask about a specific source. The example shows the real request and response shapes.
POST /api/app/memory/recall
{
"prompt": "The refund window for annual plans is"
}
200 OK
{
"tier": "certain",
"answer": "30 days",
"subject": "policies/billing-2026.md", // found from the question
"resolved": true,
"receipt": {
"source_key": "policies/billing-2026.md",
"span": [1204, 1257], // where in the store
"evidence": "Annual plans may be refunded within 30 days of purchase.",
"ingested_at": "2026-08-14T09:12:07Z",
"event_id": "8f2c…"
}
}
— or, when the evidence isn't there —
{
"tier": "gap",
"gap_reason": "no_source"
}Built for regulated operations
Enforced, not promised.
One store per customer, isolated at the database layer
Each customer gets its own memory store. Isolation is enforced by the database on every query, not by application convention. Client accounts are separated within it.
The ledger is the truth
Every addition, retraction, supersession, and erasure is an event. Replaying the ledger rebuilds a customer’s memory exactly — cached state is only a cache, encrypted with a key unique to that customer.
Curation is a human act
Retract, supersede, and redact require a signed-in person with the right role. A leaked API key cannot erase or rewrite what your organization knows.
Your data, your AI.
Everything your memory learns, it learns from you and for you. Your documents, corrections, and retractions never train a shared model, and a gap in your memory is never filled from someone else’s.
Continuous learning, by design
Memory that never stops learning.
Reasera Memory grew out of years of research into language models that learn continuously — from each event as it arrives, keeping provenance, never trading exactness for fluency. Inside the Reasera platform it holds what your organization knows and proves it, while classifiers and a re-ranker learn from your team’s corrections as they work. A gap in memory can be routed to the right model, tool, or person, and the routing itself keeps learning. The memory mechanism is patent pending.
More continuous-learning models are coming from the same research, each held to the same evidence gates. See the platform →
Watch it decline a question it shouldn’t answer.
Live, on your documents. A 30-minute walkthrough with the founder.