Demo
Same question. Same AI. One history more.
Three work situations. We ask the same model the same question, with the same instructions. One difference only: what Nawame returns from the history. The answers below are recorded outputs, copied without editing.
Mistral Large (mistral-large-latest) · recorded on 2026-10-13 · fictional data from a demo organisation
Scenario A · A replaced decision
What Nawame knew
- March 2026: Nordhost chosen to host Atlas (640 € a month).
- June 2026: Helvio replaces Nordhost, which announced +38 % at renewal, above the 700 € monthly cap.
- September 2026: migration to Helvio completed, without downtime.
- The March decision is marked "replaced": Nawame no longer serves it as current.
Without history, the AI suggests hosting providers at random. With Nawame, it answers Helvio, says Nordhost was replaced and why. The March decision, marked "replaced", is not served to it as current.
Scenario B · A constraint decided months ago
What Nawame knew
- March 2026: team agreement, nobody on support works weekends anymore.
- Consequence decided the same day: launches only from Tuesday to Thursday.
- April 2026: customer commitment to answer within 4 business hours.
Without history, the AI plans weekend on-call support. With Nawame, it recalls the March agreement and suggests moving the launch: a constraint from seven months ago changes the answer.
Scenario C · A Method and a history
What Nawame knew
- Lesson from Nordhost: no signature without a written cap on increases (5 % a year at most).
- Required: reversibility (full export and 30 days of support), support in French.
- Fallback: stay on the current monthly plan at 590 € with no commitment.
- Offer received from Helvio: 545 € a month, fixed for 12 months then indexed with no cap.
The AI alone gives a generic list. The "Prepare an important meeting" Method alone gives the right structure, but without history it can only leave blanks to fill. Method and memory together: the same structure, filled with the real offer received, the real thresholds and the real fallback.
What the demo shows, and what it does not
An honest proof, not a promise.
With the same question and the same model, the context returned by Nawame changes the answer: the decision in force rather than an old one, a forgotten constraint that comes back, a Method applied to real figures rather than empty boxes.
Its limits
Fictional scenarios, built to illustrate, and one model per page. Nawame does not make an AI infallible: it gives it back what was kept. If nothing was kept, there is nothing to give back. The best test is still one of your real topics.
Beyond three examples
On our bench of 144 scenarios (two model families, blind judging by a third family), the facts of the file are reused correctly in 92 % of answers with memory, against 20 % without, and a replaced decision is almost never served as current. A Method alone does better than a good prompt, modestly; with memory, the gap becomes clear. Some models invent more figures when a Method asks them to fill boxes: memory is what fills them with real facts.
Protocol. Same model (Mistral Large (mistral-large-latest)), same question, same work instructions (an English version of the French ones):
You are the work assistant of Sophie, director of Studio Lumen, a 12-person web agency. Answer in English, directly and actionably, in at most 120 words. If some information is missing, say so in one sentence rather than inventing it.
"With" condition: the real answers of Nawame's reading tools (gathering by topic, then reading the memory) are given to the model as tool results, exactly as a connected AI receives them. "Method" condition: the Method is loaded the way an AI that activates it does. No answer is rewritten; first run, temperature 0.3, fixed seed; two runs per condition recorded; the memory itself stays in French; only typographic change: long dashes shown as a comma or a hyphen; two redactions: the real anchor code of the demo organisation is replaced by NW-XXXXXX, and an email address invented by the model is moved to a reserved domain (.example).
Run the test with your real question.
Connect your AI in two minutes, keep two or three decisions from a real project, then ask your question again in a new conversation, or in another AI.
free · 10 memories included · Methods included · hosted in Europe