Memory is the missing infrastructure
Every era of computing got a memory abstraction — registers, RAM, virtual memory, caches. Agents got a fixed window and a prayer. We're fixing that.
Our story
Context Controller started as an internal tool. We were running long-horizon coding agents in production and kept watching them die the same way: brilliant for an hour, then the window filled, the front got truncated, and the agent forgot its own plan.
We tried everything — naive summarization, bigger models, RAG bolt-ons. Each fixed one failure mode and created two more. So we built the thing we actually wanted: an operating system for context. Priorities like process priorities. Compaction like garbage collection. Recall like swap.
Today that internal tool is a platform, and the agents running on it hold week-long sessions without losing the thread.
What we believe
Eviction should be a policy, not an accident
Truncation from the front is the context equivalent of pulling random RAM sticks. Memory pressure deserves deterministic, auditable rules.
Forgetting should be reversible
Evicted content isn't garbage — it's cold storage. An agent should be able to remember anything it once knew, on demand.
Memory is measurable
You can't improve what you can't see. Every byte of context should be accounted for, attributed, and tunable.
Small team, deep in the weeds
Maya Kessler
Co-founder & CEO. Previously led agent infrastructure at a Fortune 100 lab.
Devon Okafor
Co-founder & CTO. Built distributed caches serving 2B requests/day.
Rin Sato
Head of Research. Works on compaction algorithms and recall ranking.
Alex Lindqvist
Founding engineer. Owns the proxy, SDKs, and observability stack.
Team profiles are illustrative.
Come build the memory layer with us
We're hiring engineers who think about tokens the way kernel hackers think about pages.