Modern systems are extraordinary at storing what happened. We work on the harder problem of how experience becomes lasting capability: memory that compounds, understanding that keeps improving, and judgment that can be traced and trusted.
Organizations are humanity's most powerful mechanism for accumulating capability, yet the systems they run on preserve artifacts, not understanding. The reasoning behind a decision, the judgment of an expert, the lesson from an outcome: most of it is transient, fragmented, or lost.
Thinkers Lab is a research lab dedicated to changing that. We study the science of long-horizon memory, continual learning, and provenance: the foundations a system needs to help an organization genuinely learn over time, across its people and its increasingly capable AI agents.
Each of these is an active area of investigation. We build in the open where we can, share benchmarks, and publish what we learn.
How can a system retain, consolidate, and recall understanding across years and thousands of interactions without decay? We study memory that compounds instead of degrading, and what it should keep at full fidelity, abstract, or forget under real-world constraints.
How should a system update its understanding over time without forgetting what it already knew? We investigate parameter-efficient and reinforcement-based updates that let knowledge accumulate continuously rather than resetting with each new model.
How do we connect decisions and actions to their delayed, real-world outcomes, so a system can learn which judgments actually proved right, not just what was said? We work on credit assignment and reward modeling over long horizons.
Who said what, and where did it come from? We build robust attribution and provenance across text, audio, and vision, so every piece of understanding remains traceable to its source, the foundation of trustworthy, auditable memory.
How can memory and learning run efficiently and privately wherever an organization's data must live? We study compression, consolidation under resource constraints, and distributed, privacy-preserving learning, so understanding can be built without moving sensitive data to where it doesn't belong.
We care as much about how we evaluate progress as about the methods themselves. The hard problems in this field are easy to claim and difficult to prove.
We build the evaluations that measure long-horizon retention, attribution, and outcome prediction, so progress is real, not asserted.
Trust is foundational. We treat traceability and privacy as design constraints, not afterthoughts.
We share datasets, benchmarks, and findings with the research community wherever we can.
If you're excited by long-horizon memory, continual learning, multimodal attribution, or efficient and private ML, we'd love to hear from you, including scoped internship projects you can own end to end.