A company produces meaning every day — in meetings, in calls, in messages. Almost all of it disappears. Tessbrain turns what happens into memory with structure that answers questions and comes back as work.
This is not note-taking and not document search. It is a model of the company: people, projects, clients, decisions, tasks and metrics — and the relationships between them. It grows out of conversation on its own, with no forms to fill in.
Three things break it independently, and every one of them is familiar.
A person leaves and the context leaves with them. Not the files — files stay. What leaves is the knowledge of why it was decided that way, what has already been tried, and how it turned out.
"I think we decided this already." There was a decision — but when, by whom and on what grounds, nobody remembers. Searching the chats does not help: they contain correspondence, not decisions.
An ordinary AI cannot be trusted in a board meeting. It answers confidently and without a source. An answer you cannot verify is an answer you cannot present.
Extraction is not summarisation. About thirty specialised extractors each pull their own layer out of a transcript, and every fact gains what an ordinary summary lacks: a source, a date and a confidence.
These facts do not then sit in a list — they become nodes and edges. A task is linked to a person and a project, a decision to the discussion where it was made, a metric to the source of its number. It is the relationships that make it possible to answer questions a list cannot.
A project budget, a person's role, a task deadline are kept as a chain of dated versions. A new value does not erase the old one — it marks it superseded, with the date and the source meeting. So the system answers not only "what is it now" but "what was it, when did it change, and why".
Semantic search works like this: text becomes a vector, the question becomes a vector, and the nearest ones are returned. It works beautifully as long as the answer sits in one place. But such search has a proven mathematical ceiling (work by the Google DeepMind team, ICLR 2026): the number of document combinations a single vector can return is bounded by its dimensionality. Neither better data nor a bigger model cures that.
The easiest way to see it is an enumerative question. Try it yourself.
The point is not that the vector is "bad". The point is that the question demands completeness, while semantic proximity delivers similarity. That is why there are three channels at once — meaning, exact terms and relationships — and enumerative questions weight the third.
This is not an argument but a measurement: on the class of questions that span many conversations, ordinary top-fragment selection gives 0.23, while feeding whole conversations that contain hits gives 0.567. measured LongMemEval run, 30 questions.
The most underrated part. The system remembers not only what was decided but how it turned out — and next time says so in advance.
The key subtlety is where confidence comes from. Ten discussions of a topic whose fate is unknown produce low confidence, not high. Otherwise the number would mean "we talked about this a lot" while being read as "we verified this".
Memory that forgets nothing bankrupts its owner: what kills a long-lived system is not its starting size but its growth slope. So the system measures its own volume and growth. But forgetting here is typed: contextual links — mentioned near each other — weaken, while structural facts — decided, assigned, changed role — never decay. Otherwise forgetting would eat the org chart and the decision history.
Access control is not an interface filter. What must stay invisible is cut before the material reaches the model — which is why the model cannot let slip something it never saw. That is a stronger guarantee than asking an assistant not to talk about salaries.
What accumulates does not sit as dead weight. It comes back in seven forms, and almost all of them start without a human.
The name is literal: while a person walks from the meeting room to their desk, drafts are already waiting matched to their role in that meeting — a strategic summary for the founder, task cards for the project lead, a technical specification for the developer, a draft letter for sales. The role is determined by how the person participated, not by a line in the staffing table. You sit down not to an empty screen but to a draft that needs correcting.
The one section where "we are better" can be said with a number — and where every number has its price next to it. Runs on public sets and on somebody else's open harness; results live in the repository together with the code that produced them.
Honestly about sample sizes: 20 to 200 questions. This is real measurement, but not statistics — a difference of two or three hundredths on thirty questions cannot be taken seriously, and we do not take it so.
This is part of the product, not fine print: a product that promises more than it delivers loses trust at the first demo.
Nobody assigns a classification automatically. All facts start at "internal"; a higher level has to be set by hand. The scenario "the deal discussion will not surface for an intern" requires that somebody marked that discussion.
How much quality a closed perimeter costs has not been measured. All our numbers come from cloud models. The tool for measuring on your own model is ready; the measurement itself never happened.
There is no agent-to-agent negotiation. An external executor takes a task and returns a result under machine-checked acceptance — that is a request/response scheme, not a dialogue about how to do the work.
Company experience accumulates from the moment you switch it on. Outcomes cannot be reconstructed retroactively: for an archive already loaded, the layer appears only on reprocessing.