About
About X-Thought Labs
An independent research lab working on memory and continuity for long-lived LLM systems.
Most LLM systems today are built around a single interaction. We are interested in systems meant to work for months or years: systems that must remember what happened, know when it was known, keep their commitments and change their minds without erasing the path that led there.
Our work moves from questions to principles to working prototypes. Each prototype is built in small stages, and a stage is closed only with automated tests on the target machine and a written report. Properties that matter are registered as proof obligations and taken into formal verification.
We publish the problems we study, the principles we work by and our progress. Architectures and implementation details remain private while they are developed and reviewed.
- Focus
- Memory and continuity for long-lived LLM systems
- Current stage
- Memory engine — the scale stage (G2)
- Method
- Staged prototypes, automated tests, formal verification
- Models
- Work is designed to be independent of any single model or provider
- Public since
- August 2026
How we work
The principles behind every stage of our research.
Research before building
We review scientific literature first, then existing systems, then community sources — and compare what they say before deciding.
Convergence over single sources
A conclusion supported by independent sources carries the most weight. Disagreements are recorded, not smoothed over.
The model proposes, the memory verifies
Text written by a model does not become a fact, a decision or an obligation without passing checked rules.
Proof before training
Rules are proven before any model is adapted to follow them.
History is not rewritten
Corrections are added as new records, so what was known at a given time stays answerable.
Evidence that can be re-run
Every stage leaves its tests, results and reasoning behind, so conclusions can be checked again.