Continual-learning AI for enterprise
A model that remembers your data,
and knows when it doesn’t know.
Remembers what you tell it, for as long as you want.
Says “I don’t know” instead of guessing.
Every answer comes with a receipt: the source, the line, the date.
Tell it once it was wrong. It never says it again.
Tell it to forget, and it is gone everywhere.
Reasera Memory is a model with memory: it answers from what your organization knows, shows its evidence, and learns from every fact and correction your team gives it, with no retraining. It is the first of a family of continual-learning models built the same way.
Built on continual-learning research · Patent pending · Per-client isolation · Every answer traceable, every change on the record
The problem
Today’s AI is fluent and forgets.
Language models answer everything with the same confidence, whether they know or not. Retrieval hands them the right document and then lets them say anything anyway. Fine-tuning teaches one fact and damages others, and when something turns out to be wrong, nothing takes it back.
Receipts on every answer
which source, which passage, when it was added
Declines instead of guesses
a certified “no record” instead of an invented answer
Corrections that stick
retract a claim once and reworded copies are refused too
Why not RAG?
Same documents. Same questions. Different discipline.
We put the same documents and the same questions on both sides. One side uses standard RAG: retrieval feeding an 8-billion-parameter language model, the usual way to connect a model to your documents. The other is Reasera Memory, a model with memory, which answers from the memory itself. Retrieval handed RAG the right document almost every time.
Retrieval was not the problem. What the model is allowed to say next is.
When it answers, how often is it right?
Questions a small language model cannot answer on its own, with the answer present in the record. Reasera gives 34% of them a certain answer, another 29% a likely or possible answer (right 89% of the time), and marks the rest attested: stated by the source, no certainty level yet. Standard RAG answers every question. (3,903 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Withdrawn claims that stayed withdrawn after the same claim came back in different words and the question was asked again.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: 0 of 19,002 checks. (3,167 claims, two rewordings each)
Does it admit when it has nothing?
Questions where the subject’s record was in memory but did not contain the answer. Reasera declined 170 of 176 at the certain level. RAG, as normally configured, does not decline: it answered all 176 from the model’s own recollection, which is unverifiable and was wrong 20% of the time. Told it could say “unknown”, the same model still answered 168. (176 questions)
Evaluation sets drawn from all 21,919 CounterFact records, one per subject, with Wikipedia summaries as the records. Memory holds 9,792 of them; the 3,903 held-out questions are the ones scored. 176 questions whose record lacks the answer; 1,000 asked against the wrong record, on which Reasera gave 0 certain answers; 3,167 retractions with two rewordings each, 6,334 trials. Reasera’s answers come from memory itself: no language model writes them. Standard RAG = dense embedding retrieval, top document from a 10,650-document pool, Llama 3.1 8B Instruct generating freely; retrieval found the right document 98% of the time; the same word-level scoring rule for every arm. The 8B model alone, without the document, gets 23%. Scored against the dataset’s single answer; counting an answer as right when that answer is a member of the form it was served in (a list, an alternative, or a nesting such as city, state, country), the certain level is right 98.5% of the time and standard RAG 34%. The 29 certain-level misses are kept with the results and reviewed by hand. A fixed sample of these sets gates every release.
Results are from Reasera’s fixed evaluation sets, 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 memory works
Three auditable steps.
Add your sources
Documents, transcripts, tickets, knowledge bases. Every addition is an event on a ledger, so the state of your memory at any moment can be rebuilt from it.
Ask, and get a receipt
Each answer carries a certainty level and a receipt: which source, which passage, when it was added. When the evidence isn’t there you get a reason instead of a guess: no source, insufficient evidence, conflicting sources, or retracted.
Curate as you learn
Retract a claim that turned out wrong and it stays retracted, even when the same claim shows up again in different words. Erase what must not exist and the content leaves the ledger too. Both take a person. An API key cannot do either.
The platform
Lifelong learning, on your data.
Reasera keeps your AI learning after it goes live, from your documents, your team’s corrections, and every interaction. It needs no retraining and loses nothing it already knew. Memory holds what your organization knows and proves it. Models that route, classify, and rank learn from every correction as your team works. Everything they learn stays yours: on a ledger, isolated per client, replayable, erasable.
Recall and verify
Memory
Evidence-bound answers with receipts, certified absence, retraction that survives rewording, and erasure that is real.
Product page →Route and adapt
Classifiers
Ticket triage, document sorting, intent detection. Add a category with a single example and it is live without a retraining run. Every correction updates the live model, with no retraining and no data-science team.
Select
Re-ranker
The step that decides what reaches the user or the LLM. Improves with every rating, with no fine-tuning and no retraining project.
How the pieces work together
When memory cannot certify an answer, that gap tells you where to route the question. Route it to a specialist model, a tool, or a person, using a classifier that learns new routes as you add them. New client, new product line, new tool: the platform adapts as requests arrive, with no retraining project. The family keeps growing: more continual-learning models from the same research, each held to the same evidence gates.
Built for regulated operations
Isolation, audit, and erasure are enforced by the system.
Per-client isolation
Each customer gets its own isolated store for memory and models. Your clients' accounts are separated within it.
An auditable ledger
Every addition, correction, retraction, and erasure is recorded. Any past state can be rebuilt from the ledger, minus what you have erased.
Controlled model deployments
Every model deploys the way you choose. Hold new learning at a checkpoint for validation, or let it go live with every correction.
Curation needs a person
Retracting or erasing memory takes a person with the right role. API keys cannot do it, by design.
Your data stays yours
Everything your AI learns, it learns from your own documents and corrections, and none of it trains a shared model.
Fits your stack
HTTP API, memory-scoped API keys, webhooks, and an MCP server so agents and copilots can use memory directly.
Where Reasera fits
When your AI needs to move as fast as your business.
AI agents and copilots
Give agents a memory that cites its sources and says so when it has none. Available as an MCP server and an HTTP API, so it works with the agent framework you already use.
Customer operations
One platform, many clients, each isolated. Ticket routing that learns from every correction, and a private memory for each client.
Regulated knowledge
Finance, healthcare, legal, compliance: answers must be traceable and mistakes must be retractable. Cite-or-decline is the default, and the ledger is the audit trail.
Search and RAG pipelines
Keep your retrieval; add discipline. The re-ranker decides what reaches the model, and memory decides whether the model may answer at all.
See it on your own data.
A 30-minute walkthrough with the founder. Bring a question your AI gets wrong.