16 November 2026 · 4 min read
Memory you can trust: how Anabel remembers, forgets and keeps secrets
Our white paper on AI memory for companies: how Anabel was tested on remembering, forgetting and keeping secrets, and what the results show.
Download the white paper (PDF)16 pages. Second edition, revised 16 November, with the 12 November results. Free, no sign-up.
Language models are now very good at writing and reasoning. What they lack is memory. Every conversation starts from nothing, unless something outside the model decides what to remind it of.
For one person chatting with an assistant, a simple memory is enough. A company is different. A company assistant hears from the chief executive, finance, operations, clients and suppliers, over email, calls, meetings, documents and chat, every day. Our white paper sets out how we built Anabel's memory for that job, and how we tested it.
Three hard things at once
An assistant that works for a whole company has to do three things together, and they pull against each other.
- Remember the right thing, including what was said weeks ago, on another channel, by another person, and the latest value when a figure has changed.
- Forget on request. When someone says forget that, it has to stay out of every later answer, even when someone tries to get it back.
- Keep people's information apart. Finance's figures must not reach someone not cleared for them, and one company's information must never reach another.
A memory that recalls more has more to leak. A memory that refuses too readily is safe but useless. The only honest test scores all three at once, so strength in one cannot hide weakness in another. Most memory systems are built and measured on the first alone. Anabel was built for all three, and we tested her on all three.
What the tests found
We used two tests. The first is GateMem, a public benchmark published in June 2026 by independent researchers, which scores remembering, forgetting and keeping secrets together. The second is our own company-wide test on a full company setup with email, meetings, documents, notes and customer records.
- The Anabel Base scored above the memory and retrieval baselines in the GateMem paper in all four areas: medical 75.9, office 72.6, education 46.1 and household 56.4, out of 100. These come from our second reported measurement of the Base, on 12 November, after a round of memory improvements. The first, on 7 and 8 November, averaged 57.1.
- The four-area average was 62.8, about 54% above the paper's best single fixed setup at 40.85. Against the best score in each area taken separately, the comparison is about 45%, with a baseline average of 43.35.
- On office work, the Base scored 72.6 against 58.5 for the paper's best memory and retrieval baseline. Giving the AI every word of the conversation scored 66.7 in our earlier run on the same model. Whole conversation still scored slightly higher overall across the four areas: 64.8 against 62.8.
- On GateMem, when asked to forget something, the Base kept it out of answers in 94.6% to 97.7% of checks, depending on the area, on tests built to coax it out.
- In our separate earlier company-wide engine test, 97.1% of 205 untouched recall questions were answered correctly, and 98.0% of all 307 recall answers named the right source. This test has not been repeated on the Base.
- In that company-wide test, none of 161 attempts to reach another person's or another company's information got through. This is an observed test result, not a guarantee of zero leaks.
Why the walls hold
The company-wide test checks the storage boundaries: each person's Anabel keeps its own memory, and shared information is copied only into memories allowed to hold it. GateMem deliberately uses a shared memory pool, so its privacy scores test a different problem. The white paper reports those scores separately.
Time matters too. Every fact carries when it was true, so Anabel knows which version of a price, a date or a decision is current, and can tell you what it was before.
How we ran the tests, and what is in the paper
All tests used synthetic data, with no client or real company data. We ran GateMem ourselves on Claude Sonnet 5.5 with the official evaluator and judge settings; the paper's published figures use other answering models. These are comparisons across complete setups, not a controlled comparison of memory systems on one model. No outside party has verified our runs.
The Base measurements come from memory built on Base code on 11 and 12 November 2026 and scored on 12 November, with one reported run per area. For the test, a setting that gives memory the asker's recent thread was switched on; it is on by default in the released Base since 15 November. The same memory code, with later fixes to forgetting, now answers WhatsApp chat on a trial installation of the Base. That is separate from these benchmark results, which are not live client results. The white paper gives the method, all scores, revisions and limits.
If you are weighing up AI for your company, it is worth reading before any vendor conversation, ours included.
Download the white paper (PDF)16 pages. Second edition, revised 16 November, with the 12 November results. Free, no sign-up.
Want to see it in your company?
Tell us how your business runs and we will show you what Anabel would do in your first week.
Talk to us