The thinking core

Eighteen agents each remember. The company cannot ask.

Brainix collects what every agent remembers, distils it tier by tier, and regroups it into company-level themes. The model runs on your premises — local, or a cloud model you choose; every conclusion carries the source text behind it.

Talk to us about Brainix What you can check

Why Brainix exists

This is not a choice of tools. It is four facts that hold at the same time.

Eighteen agents keep their own memory, and nobody joins it up. Answering "what are we doing, and where is it stuck?" means asking each of them in turn.
103,369 raw files, about 55 GB, measured. No person and no single model call can read that.
Cloud memory services require company conversations and credentials to leave the building. Clients in Hong Kong do not accept that.
Management wants the reason behind a judgement, not an answer that cannot be traced.

Four layers: ingest, distil, feed back, surface

All four have to hold at once for memory to be usable. Each layer's state is stated plainly; what is not built is not dressed up as a demo.

Ingest

Collects each agent's five memory tiers: raw, distilled, mid-term, project and big picture. State: runs daily.

Distil

Summarises each agent, regroups with a sliding window, then clusters across agents into company themes. State: runs daily.

Feed back

Writes back to the company project tier and returns results to each agent. State: writes pass; delivery to agents still being built.

Surface

Query interface, weekly report, decision brief. State: being built.

What already runs, in numbers

Measured in November 2026. Every figure can be recomputed on the spot.

103,369 raw files, about 55 GB, held across eighteen agents.

A five-tier memory standard — raw, distilled, mid-term, project, big picture — in force across the team, each tier measured.

11 company-level project files whose index matches the real files line by line.

Daily collection and daily regrouping run unattended; one pass regrouped 41 work lines into 5 themes in 9 minutes.

Six agents have received their per-agent feedback file; the rest have empty or unmounted memory, so there is nothing to analyse yet.

What you can check

Bring a question you cannot answer today. We run it on your own memory.

Any summary can be traced back to the source text it was condensed from.

File counts and volume can be recomputed on request.

Every work line states which agent it came from and on which day.

Every model call passes through the Brainix gate; the log can be inspected on request.

Layers that are unfinished are labelled unfinished — no demo screens standing in for them.

Bring a question you cannot answer today

What are we doing, who is doing it, and where is it stuck? We will run it on your own memory, so you see the answer and the source text behind it.

Talk to us about Brainix on your premises