Focus AI Brain

Focus AI BrainA company memory you can act on.

Your team decides things in meetings and forgets them by Friday. The Brain turns your meetings, decisions and projects into a company memory that checks every fact before it counts, and answers with its source, or tells you it does not know. Built for European companies that run on meetings and project tools.

Request a demo Try the demo

Running in production today

Gate  /  candidate 0418
  1. Ingested
    Weekly sync, 15 Sep · transcript 00:23:41
    “From now on, every discovery call runs on the standard scoping template. No more free-form notes.”
  2. Distilled
    Decision: all discovery calls use the standard scoping template. Replaces free-form discovery notes.
  3. Judged
    • Verifiable from its source Pass
    • Not already known Pass
    • Replaces an approved note 1 note
    • Still true in six months Pass
    • Needs a person (roles, contradictions, supersedes, deletions) Yes, supersede
  4. Verdict
    ApprovedSupersede of the older discovery note confirmed by a person
  5. Stored
    Note
    decisions/discovery-scoping-template
    Source
    weekly-sync-2026-09-15, 00:23:41
    Approved by
    gate judge + person
    Status
    active

Example data. The same process runs in production.

Your company forgets more than it learns

  • MeetingsDecided on Monday, forgotten by Friday
  • Project toolsWritten down once, never found again
  • PeopleWalks out the door when they do

Plugging an AI assistant into that does not fix it. The assistant retrieves stale and unverified notes with total confidence, so you get bad answers, delivered faster. The Brain fixes the input first.

Follow one decision through the Brain

This is the path candidate 0418 took through the six layers of the Brain, from a meeting to a cited answer. The model that distills never judges, the model that judges never writes, and code does the writing.

  1. Ingestion

    The weekly sync recording arrives with its metadata: who, when, subject.

    weekly sync · 15 Sep

  2. Processing

    An adapter normalizes the provider’s transcript into clean text with metadata and a hash. The original stays in the archive.

    transcript 00:23:41

  3. Digestion

    Triage flags minute 23, a model distills one decision, a judge checks it. It replaces an approved note, so a person confirms.

    approved · supersede confirmed

  4. Storage

    Code writes it to the knowledge base through the single staging area, versioned in git, with a receipt, source, date, approver and the hash of the meeting it came from.

    decisions/discovery-scoping-template

  5. Indexing

    It is indexed by meaning and by structure: which meeting, which people, which project.

    semantic + entities

  6. Retrieval

    It reaches the next AI session that touches discovery without anyone asking, and comes back as the cited answer when someone asks.

    cited answer

Anything that changes a person’s role, contradicts approved knowledge, or supersedes, merges or deletes a note is never decided by a model. It waits for a person.

Try it: run a meeting line through the gate

Click a line that was said in a meeting. It runs through the gate on its own: four lines, four different outcomes. Then ask the memory a question and see how it answers, and when it refuses to.

Gate  /  ready
  1. Distilled
    Click a meeting line on the left.
  2. Judged
    • Verdict
    Your question

    Choose one of the questions to see the answer and its sources.

    Example data. Runs in your browser.

    Six layers, two spines, two loops

    Under the hood, for technical reviewers. One rule runs through all of it: deterministic work is done by code, and a model is used only where judgment is needed. Whoever distills does not judge, whoever judges does not write, and whoever writes is code.

    GOVERNANCE · WRITE PATHRELIABILITYSPINES · CROSS ALL LAYERS.USAGE SIGNAL · BACK TO DIGESTION01 INGESTIONevery source is a capturer02 PROCESSINGone adapter per format03 DIGESTIONworth a look? what enters?04 STORAGEmarkdown + git · archive · registry05 INDEXINGby meaning and by structure06 RETRIEVALpush · pull · browse · chat · rerankknowledge flows down · context rises to sessions · the usage signal returns 01 INGESTIONevery source is a capturer02 PROCESSINGone adapter per format03 DIGESTIONworth a look? what enters?04 STORAGEmarkdown + git · archive · registry05 INDEXINGby meaning and by structure06 RETRIEVALpush · pull · browse · chat · rerankSPINESgovernance (write path) · reliabilitycross the six layersTWO LOOPScontext rises to sessionsusage signal returns to digestion

    1. 1Ingestion

      Every source is a capturer. Meetings from any recorder, email threads, project tools such as ClickUp (tasks, comments and chat), WhatsApp Business in listen-only mode, and AI work sessions. Each item arrives with its metadata (who, when, with whom, subject), so what clearly does not matter is discarded before any model is paid for. Conversations with the chat assistant are a source too, and a mandatory secret scan runs before any external model sees the content.

      • configurable providers
      • metadata pre-filter
      • provenance per source

    2. 2Processing

      You do not filter audio. You filter the text that came out of it. One adapter per format (transcript, email, chat, PDF, image, audio, code), all with the same output: normalized text, metadata, a hash and a reference to the original, which stays intact in the archive. Nothing is decided here.

      • format adapters
      • text + metadata + hash
      • original preserved

    3. 3Digestion

      Two questions with different owners. Triage asks “is it worth a look?”: high volume, a cheap model, and anything it passes over stays marked unreviewed, never healthy. Distillation and judgment ask “what exactly goes in?”: a stronger model distills candidates and a separate judge approves, edits, discards or escalates. A person’s role, a contradiction, a delete, a merge or a supersede always goes to a person. Digestion trims, it does not discard: the rest stays in the archive.

      • triage
      • distiller + judge
      • human escalation

    4. 4Storage

      Three shelves, one database. The knowledge base stays as markdown in git, with status, provenance and cheap revert. The archive keeps everything that came in, raw, addressed by the hash of its content. The registry links entities: person, company, meeting, task, project. PostgreSQL with pgvector holds the archive, the registry, the ledger of write receipts and an indexed copy of the knowledge base, always rebuilt from the markdown, never the other way round. Every note points to the hash of the archive item it came from.

      • markdown + git
      • content-hashed archive
      • entity registry
      • receipt ledger
      • PostgreSQL + pgvector

    5. 5Indexing

      Two indexes: by meaning and by structure. A multilingual semantic index finds what is similar to the question. A structural index knows that Tuesday’s meeting had three people and produced two decisions. The archive has its own semantic index, separate from the knowledge base, so raw material never contaminates trusted search.

      • semantic index
      • entities and links
      • separate archive index

    6. 6Retrieval

      Retrieval is the product. Push: relevant context is injected into every AI session automatically. Pull: search, API and MCP. Browse: topic maps and links. Conversation: a chat assistant with the brain as context, by text, audio or photo, with identity locked, every command logged and confirmation before anything irreversible. At the door, a reranker reorders the top results and sets how much context each prompt gets. A question about a specific item is answered only with that item open and cited; otherwise the answer is “I couldn’t find it”, never a guess.

      • push / pull / browse / chat
      • rerank
      • cites or abstains

    Spine 1: governance and write path

    Not a layer, a column through all of them. Nothing writes to the knowledge base except through a single staging area, every writer leaves a receipt, and every change can be reverted with git.

    Spine 2: reliability

    Every note says how much to trust it (draft, active, stale, superseded), unreviewed material stays marked unreviewed, and the system watches itself. When something breaks, a stronger model diagnoses it read only and applies a reversible fix, within cost and attempt limits. A person is called only if the fix failed, the action is irreversible, or the decision is about content.

    Two loops

    What enters the knowledge base reaches the next AI sessions without anyone asking. The Brain also logs what it injected and what was consulted, and when someone corrects the AI right after it used a note, that signal returns to digestion, so the brain also learns from what it delivered.

    AI assistants that report to a person

    Ready-made assistants sit on top of Retrieval and take over repetitive work, each with a single job. They read the memory like any other client, a person approves what they produce, and anything new they create comes back in as a source, through the same digestion, the same gate and the same write path. No assistant writes to the knowledge base directly.

    Meeting debrief

    Turns a recording into decisions, owners and action items.

    recordingdebrief

    Discovery draft

    Prepares the discovery document from the calls and what the company already knows.

    calls + memorydraft

    Scoping assistant

    Matches a new problem against solutions already built, so the team starts from proven work instead of a blank page.

    problemmatched templates

    Delivery QA

    Checks a deliverable against the agreed scope before it reaches the client.

    deliverablereview

    New assistants are added as your workflows need them, each on the same memory and the same gate.

    Built for trust, not for volume

    Most AI knowledge tools measure how much they store. The Brain measures how much of what it stores you can rely on. Here is what that means for candidate 0418.

    History  /  discovery
    1. Marprocess/discovery-notesActive
    2. 15 Sepdecisions/discovery-scoping-templateDraft
    3. 15 Sepdecisions/discovery-scoping-templateapproved by the gate judge, supersede confirmed by a personActive
    4. 15 Sepprocess/discovery-notesSuperseded

    Illustrative data

    • Nothing enters unjudged

      Every candidate faces an independent review before it reaches the company memory, and the rules are biased against entry. Doubt means reject or escalate, never approve.

    • Every fact has a trail

      Source, date and approver travel with each note, so you always know where something came from and who let it in.

    • Knowledge ages honestly

      Notes move from draft to active, then to stale or superseded. An outdated note is replaced and points to its successor. Nothing is silently deleted.

    • Private by design

      Raw recordings never enter the knowledge base, and every answer is served through a scoped view per role and per project. GDPR is part of the architecture, not an afterthought.

    In production

    The Brain runs in production on a daily schedule, autonomously, with a person stepping in only on escalations.

    • A leading real estate company in Luxembourg runs the Brain as a pilot.
    • Focus AI already works with leading Luxembourg companies in real estate and fast food.

    Give your company a memory it can trust

    Tell us how your team works and we will show you the Brain on your own material.

    Request a demo