Memory

The problem

What an AI system keeps from its work is often hard to inspect, hard to trust and hard to recover exactly.

Stored knowledge can be altered without anyone noticing, recalled with more confidence than it has earned, or lost as the system around it changes.

For a model to be relied on in long work, its memory has to last, return exactly what it was given, and be trusted only as far as it has earned.

What we set out to do

The work brings four things together: knowledge that can be recovered exactly, altered knowledge refused rather than used, knowledge that lasts beyond the setting that made it, and confidence tied to what the work has actually shown.

Holding all four at once is what makes the problem hard.

What we found

Information came back complete

Ouah recovered the information it held completely, and took in new information without weakening that result.

Altered knowledge was refused

Ariadne refused knowledge that had been altered, and recovery without authorisation failed while authorised use carried on.

Knowledge stayed exact beyond where it was made

Ariadne recovered knowledge exactly after it left the setting that made it.

Memory that recovers and renews

Athena recovered several distinct stored memories exactly, including when the cue at recall differed from the original encounter, and renewed memories whose recall had deteriorated.

What it is for

A model that keeps and recovers what it has learned can be relied on in longer work, without silently losing what it was given.

Exact memory also makes AI work easier to review. When knowledge comes back exactly and altered knowledge is refused, what a model holds can be checked against what it was originally given.

Where it stands

Memory runs through several Forlais Research programmes. Ouah established dependable custody of information. Ariadne showed that knowledge stays exact beyond the setting that made it. Athena built the foundation for memory that lasts, recovers and renews, and the work continues there: Dynamic Memory, in active research within Athena, has its own page.