Lab

Why we publish Nawame’s limitations too

Short answer

We publish Nawame’s limitations because a memory whose weaknesses are hidden invites too much trust. Long-term memory is a hard problem: false precedents, noise, dilution and the passing of time. We measure, we improve, and we do not claim to have solved it.

A glass disc with a fine hairline flaw revealed by amber light, surrounded by measurement rings and tick marks.

On most product sites you only read what works. Here, part of the page is about what does not work yet, or not every time. This is not a pose. For a tool whose job is to hand context back to you, knowing where it can be wrong is part of the manual.

Why publish the limits of a memory tool

A memory is rarely judged on its successes, because you do not notice them. It is judged on the moment it confidently brings back something false. If nobody told you that could happen, you granted it a trust it had not earned.

Publishing the limits is a way to set trust at the right level: enough to use it, not so much that you stop checking what matters. It is also a slightly uncomfortable commercial bet, since some readers prefer crisp promises. We would rather you choose Nawame knowing its flaws than leave after discovering them.

That said, for many simple uses, an assistant’s native memory, a well-kept document or a careful prompt is enough. We have no reason to pretend otherwise.

Why long-term memory is a hard problem

Long-term memory is the ability to keep context for weeks or months and find it again at the right moment, instead of lasting only for one conversation. It is hard for reasons humans know well:

  • Time changes what is true. What was true in February may not be in August, and an accurate memory can become misleading without changing a single word.
  • Noise builds up. The more you keep, the more the useful memories drown among trivia. Keeping less often means understanding better, which is why not writing is a feature.
  • Resemblances deceive. Two situations that look alike on the surface can have nothing to do with each other. We devoted an article to the false precedent in semantic search.
  • Someone has to choose what the AI is given. Too much dilutes what matters, too little leaves out a decisive element.

We do not know a definitive solution to these four points. We work to reduce them, and we would rather say so.

What we measured, and what it does not prove

We have two public results, and only two.

The first is about durability: in Lab tests, we counted 465 confirmed writes, 0 violations observed. In other words, what Nawame confirmed it had written stayed written, as far as we observed. This does not prove that no failure is possible, and it says nothing about the quality of what is retrieved afterwards.

The second is about not writing: our first synthetic protocol for not writing correctly classified 26 events out of 26. The important word is “synthetic”: we prepared those events ourselves, in a controlled setting. Real life will be more complicated. We explain this choice to publish imperfect results as well in why we also publish our imperfect benchmarks.

These results say that the foundation holds and that the idea of not keeping everything can be tested. They do not say that Nawame always finds the right memory, or that it knows what matters without ever being wrong.

Current limits

Here is what we know about our limits, as of today.

Retrieving

Nawame can miss a relevant memory, or bring back one that is not. To keep a resemblance from passing for proof, retrievals are presented as a “precedent found”, a “possible connection” or “no convincing precedent”. This vocabulary helps calibrate trust, it does not prevent every error. And found does not mean true: a memory keeps its provenance, its date and its context so that you can judge it. What degrades as volume grows is described in limits of retrieval at scale.

Selecting

Deciding what deserves to become a lasting memory is a choice that can go wrong in both directions: keeping what should not have been kept, or letting through something that would have helped. The more the history grows, the more demanding this choice becomes, along with the choice of what gets presented to the AI.

Following time

Nawame knows that one decision replaced another when it was told so or when it is clear from what was kept. It does not guess that information is outdated if nothing indicates it. Imagine that in February you noted “we keep the current supplier”, then in August you change your mind out loud without recording anything: the February decision can then resurface as if it still held. The ways a memory goes stale are reviewed in how to tell that a memory has become obsolete.

Protecting more, understanding less

Sovereign mode strengthens confidentiality: some content stays unreadable to server-side processing. The trade-off is real: capabilities that need to understand content on the server are reduced. This mode is available from the free plan, and it is not “zero compromise”. For details, see the security page.

Taking your memory with you

Export gives you back your memory, your useful history, your replacements and their provenance, in versioned formats. The internal mechanisms of selection and retrieval are not in the export: they remain Nawame technology. Imported information stays identified as imported and does not automatically become authoritative.

Deciding

Nawame provides context. Judgment and decision stay human, even when the context is retrieved well.

How we work to reduce them

The Lab studies how humans and AI lose, retrieve and use context. Its axes come down to five verbs: capture better, retrieve better, give less but better, know when to say nothing, withstand time.

Part of the work is testing what should not be found. A memory that invents a memory is more dangerous than a memory that has none, so we also check that Nawame can answer that it has nothing convincing. The reasoning is explained in why we deliberately test topics that do not exist.

We do not publish the details of our methods, our test sets or the inner workings: that is part of our technology. We publish the principles, the results we can stand behind and the limits that come with them.

What this changes for you

Treat Nawame like a colleague with a good memory who is not infallible. For information a real decision depends on, look at the date, the provenance and the reason before relying on it. Keep your reference documents where they are, and remember that export lets you leave with your context.

If this candor speaks to you, you can start free and judge for yourself, with your own topics.

Your next prompt should not have to contain your whole story.

Start building a memory your AI tools can find again when it matters.

Start free