Layer 2 · Brain & Memory
The foundation is static truth. The brain is what makes it live: retrieval over everything we’ve ever written, a conductor that reasons across it, and a memory that turns every completed task into permanent capability.
RAG — retrieval over everything
Skills, transcripts, research, client docs, prior decisions — embedded and searchable. When an agent reasons, it reasons grounded in what the company actually knows, not just what the base model was trained on.
Documents, meetings, and code — embedded and searchable with hybrid (full-text + vector) retrieval and source citations. Refreshed on a schedule so it never goes stale.
Companies, people, trials, labs, news, and grants linked as a graph, so the system answers relational questions — who connects to whom — not just keyword ones.
A model proxy routes each call to the right engine — a free local model for cheap classification, frontier models for hard reasoning, a gated council only for high-stakes calls.
The Conductor
The Conductor is the CEO-agent — a container on the Brain that reads the pulse, queries the brain, consults the directors, and proposes the next best action, which always lands in front of a human before anything runs.
Perpetual memory
Most automations forget. Ours consolidate. Memory is scoped four ways — global, client, agent, session — and a weekly memory dream (the Conductor’s tier-2 consolidation) distills raw activity into durable, namespaced memory, so the system that did the work last week is sharper this week. This is the self-learning loop the whole OS depends on.
Every meaningful action — a resolved account, a sent follow-up, a graded decision — is recorded as a scoped memory event with context.
A weekly think-loop compresses events into learned rules and summaries. Noise is dropped; signal is kept and made retrievable. We were early on this, not behind.