Verifiable memory
Memory that can show where each record came from, when it was known and who may read it — and that does not let a model turn its own text into a fact or a decision.
Independent research lab
Memory and continuity for long-lived LLM systems
We study how LLM-based systems can keep a truthful history, honour their commitments and stay coherent over time — and we test these ideas in working prototypes before we rely on them.
Focus
What we work on
Time and revision
Answering both “what is true now” and “what was known then”. Corrections are added as new records; the path between an earlier and a later view is never rewritten.
Independence from the model
A long-lived system should outlast any single model. Its history, obligations and rules belong to the system, so the model working with them can be replaced.
Latest update
Research progress
Engine foundation complete (G1)
The foundation of our memory engine passed an independent code audit and three rounds of re-checks, and the stage closed only when no critical finding remained. Every decision is registered and linked to the tests that check it.
Next: the engine at scale (G2). Large-scale measurements are under way; we will publish them once they are final.
Read the timeline- Memory pilot
- Completed
- Engine stages closed
- G0, G1
- Automated tests in a full run
- 2,600+
- Registered decisions and rules
- 480+
Research notes
Latest notes
Convergent Evidence
Before building, we review three kinds of sources in order: scientific literature, existing systems and working implementations, and community discussion.
An Honest “I Cannot Know”
Evaluating memory means measuring what a model does without it, too.
Train on Proven Rules Only
Adapting a model to a set of rules is only as sound as the rules themselves.