The hardest part of AI memory is not storage. It is restraint. When related facts are missing or conflict, a trustworthy memory system should refuse to invent a tidy story. We design for grounded composition: combine what is known, surface uncertainty, and avoid fabricated bridges between fragments.
Early memory prototypes fail in a familiar way. Ask a question that touches two real facts and one missing link, and the system produces a fluent paragraph that sounds right. The fluency is the bug.
The failure mode
Language models are trained to continue. Given partial evidence, continuation often means invention: a motive that was never stated, a timeline that was never confirmed, a causal link that only looks obvious in hindsight.
For a chat toy, that can be entertaining. For a memory layer people rely on across work, it is unacceptable. Memory that fabricates is worse than no memory, because it teaches you to trust a record that is quietly rewriting history.
What we optimize for instead
- Prefer an incomplete but honest answer over a polished guess.
- Compose related facts when the record supports the composition.
- Keep competing values visible when the record disagrees with itself.
None of that requires publishing internal scoreboards. You can feel the difference in daily use: the assistant stops “helpfully” filling gaps you never authorized it to fill.
Why we are building in public on this point
Reliability is easy to claim and hard to demonstrate without turning a marketing site into a lab notebook. So we state the standard plainly: Archilas is aimed at memory that does not fabricate answers and that combines related facts correctly. The work continues, and the standard does not get relaxed for demos.