[work] company brain
A knowledge system that organizes itself.
Capture is easy; every second brain dies at retrieval. The Exocortex is what comes after capture: a context layer with retrieval engineered per dataset, scheduled synthesis that digests the stream, and delivery that hands the right context to the right mind at the right moment. 130+ days as the working memory of a real multi-company operation.
[01] the real problem
How do you get the right sliver of accumulated knowledge to the right mind, at the right moment?
A pile of notes does not think. A bookmark inbox does not surface the idea you needed three months later. The system we built answers the question in four movements: capture without ceremony, a true context layer, synthesis on a schedule, and handoff at the right place and time.
[02] movement one
If it required remembering, it would already be dead.
Nothing in the system depends on human discipline. Scheduled jobs dump and collect: every working session writes its own summary on exit, bookmarks flow in nightly and get enriched with tags and intent, commits and journals accumulate where the system can see them. Capture is ambient.
[03] movement two
Retrieval is a portfolio of techniques, engineered per dataset.
Hybrid semantic search where meaning matters: keyword, vector embeddings, and LLM reranking, finding the note that is about the thing, whatever words it used. Plain structural search where structure is known. Taxonomy that makes whole classes of questions answerable without any search at all. Embeddings where similarity is the question.
The design decision is matching the technique to the shape of the data. Get that right and everything downstream gets cheap; get it wrong and no amount of AI compensates.
[04] movement three
Accumulation becomes knowledge because something is paid to curate it.
Raw capture rots. So the system digests on a schedule: a librarian job compresses each week's sessions, commits, and captures into a bulletin. A daily synthesis pass connects today's inputs to months-old material and writes the connections down. Long-term memory and an entity graph grow by append, one curated fact at a time. In compute, nightly, without being asked.
[05] movement four
Synthesized context is delivered, not searched for.
Every agent session opens with the current state auto-loaded: what is true right now, what changed this week, what is worth connecting today. Four different agents operate this system, and whichever one picks up a task starts warm. The human gets the same treatment: the morning brief is waiting before anyone asks. Nobody, silicon or human, starts cold.
[06] the proof
130+ days unattended. Zero cold starts. Four runtimes swapped freely, zero memory lost. ~$0.20 a day.
This is not a design document. The system has run for 130+ days as the working memory of a real multi-company operation, with four different AI agents reading and writing one shared memory. The models get replaced as better ones ship; the substrate they read from compounds daily.
This is the company brain written as a design story, and the twin of Giving Agents Hands: that craft arms the agents, this one remembers for them. Together they are what a Mothership engagement installs. We built it for ourselves first because we needed it.
[next] explore
Powered by The Engine. The twin craft: Giving Agents Hands. Hands there, memory here.