Research
Research releases
Notes and releases from the Roketus AI Engineering Studio.
Where an Eval can say “I don't know”
How we check the agent workflows we build and run: exact checks for tools, arguments, trajectory and goal state; a judge for claims in prose; a person for intent; and a third outcome, no-signal, for when nothing was measured.
A mind that sleeps without changing its brain
A frozen LLM can consolidate its experience to an external file corpus instead of to weights. On a fair, balanced benchmark the two substrates tie at this scale (n=32) — a directional proof-of-concept, not a powered claim.
The Symbolic System Is the Labeler
A method note: when a deterministic symbolic rule graduates, it emits a supervised training example for a neural layer — self-bootstrapping labels with a temporal signature. Approach, not results.
Tiered Memory for Long-Running Agents
A short note on splitting agent memory into hot, warm, and cold tiers to keep context windows small without losing recall.