Midbrain

The Problem

Agents run in short loops.

Context in. Inference out. Reset.

They recall, but they don't update.
Twice the memory, same mistakes.

Storage isn't intelligence.
Learning is.

The Missing Loop

To improve, a system has to change because of what it saw — not just remember seeing it.

Without the update step, memory is just storage.

The Architecture

We don't ship a vector store. We ship three memory substrates under a learned controller — mirroring how human memory is actually organised.

Episodic holds raw traces. Semantic consolidates them into notes and a temporal knowledge graph. Procedural captures the patterns the agent starts to reuse.

A learned controller decides — on every event — what to add, update, link, or forget. Memory is a policy, not a write.

Retrieval vs Experience

Every memory product on the market answers the same question: what did we see?

We think the question that matters is different.

Retrieval makes the agent faster at being wrong.
Experience makes it right.

SmartSearch

You don't earn the right to build the learning layer without first winning the retrieval layer. SmartSearch is ours — an index-free, structured retrieval system for agents operating over long horizons.

Where It's Going

Phase 1 — Memory infra. Best-in-class retrieval and structure. Shipping now.

Phase 2 — Experience graph. Every interaction becomes a structured trace: who, what, where, when, why, outcome.

Phase 3 — Continual learning engine. User corrections become training signal. The controller updates. The agent adapts.

We're not building better retrieval.
We're building agents that get better with use.