what time was it again?
An R&D lab studying memory as a foundational substrate for . Across language, vision, audio, and world models.
read.the.work
What we study
[1] Persistent coherence
The same system, across sessions and model upgrades, should be able to refer to a single intelligible past. We study what makes a memory layer hold its shape under churn — and where the cache-as-history pattern silently fails.
[2] One substrate, many runtimes
We are looking for the minimum representation under which agents, workers, and simulation stacks can share the same memory. The hypothesis: most “retrieval pipelines” are reinventing the same substrate poorly.
[3] Compound intelligence
Memory should compound, not flatten. We study promotion, revision, and consolidation as first-class operators — borrowed in spirit from how biological memory keeps sharpening rather than degrading with scale.
[4] Governed crossings
What crosses into a model’s context is a decision worth examining. We treat selection, budgeting, and explanation as part of the memory problem, not a layer bolted on top.
[5] Trust at machine speed
Lineage and receipts are how memory becomes legible — to operators, to evaluators, and to the system itself. We are interested in what changes when a system can inspect its own recall.
[6] Prove it, then ship it
Memory failures are slow to surface and expensive to unwind. We build the evaluation harnesses we wish existed, and publish what we learn — including when the runtime is wrong.
Research Areas
The most-studied memory in modern AI is the context window. We are asking what comes after it: a representation in which meaning, phrasing, and time live in the same space, and where what an agent learns can stack rather than dissolve at the next session boundary. The runtime is the place we test what we believe.
Open questions we are working on right now:
Temporal coherence is a memory problem in disguise. We study what it takes for a generative video model to keep character identity, scene continuity, and physical state across hundreds of frames — without drifting into a different world halfway through. Early notes argue that frame-level conditioning is not enough; some form of persistent state is required.
Long-horizon agents — in games, simulators, and eventually robotics — need to maintain and update beliefs about an environment that outlasts any single rollout. We are interested in persistent state layers that survive resets, generalize across instantiations of the same world, and stay legible enough to debug. This is where memory, planning, and identity start to look like the same problem.
Storage is cheap; selection isn’t. The interesting research lives in the choice.
