Skip to content
OnDuty

What reaches the model

A company brain is judged on one thing: when a model answers, was it holding the right few hundred words. Everything else here is upstream of that. This is the part that is usually invisible, so this page shows it working.

The same customer, three readers, three blocks

An agent answering on WhatsApp, a member reading a board and an outside assistant with a window of its own are not the same reader. Move the budget and watch what the Brain serves, and what it finds and drops.

Context block
Who is reading

Something a customer typed at the model

Budget spent959 of 1400 tokens

In the block

Who

  • - (hubspot, 2026-08-21) customer: Marcela Rivas (+57 300 ··· 2233). First seen through whatsapp. [demo-ent-marcela]

Connected to

  • - (hubspot) ordered: order 4471 [demo-ent-order-4471]
  • - (whatsapp) wrote in: conversation of 21 August [demo-ent-conv-0821]

Known (11/17)

  • - (preference, high confidence, whatsapp, 2026-08-12, seen 2x, confirmed by a person) Wants the delivery on Friday rather than Thursday. [demo-mem-marcela-friday]
  • - (commitment, high confidence, hubspot, 2026-08-05) Order 4471 was promised for the week of 18 August. [demo-mem-marcela-date]
  • - (observation, medium confidence, whatsapp, 2026-08-17, seen 3x) Answers on WhatsApp and rarely opens email, so the reply should go where she wrote from. [demo-mem-marcela-channel]
  • - (inferred, low confidence, hubspot, 2026-07-30) Probably buying to resell rather than for herself, from the quantities on the last three orders. [demo-mem-marcela-reseller]
  • - (preference, medium confidence, gmail, 2026-06-11) Asked for the invoice to carry the company name rather than her own, and for it to arrive before the goods. [demo-mem-marcela-invoice]
  • - (fact, high confidence, hubspot, 2026-06-02) The delivery address changed in June and the old one is still on two open orders. [demo-mem-marcela-address]
  • - (observation, low confidence, whatsapp, 2026-05-19, DISPUTED by a person) Was told once that a discount might be possible on a larger order, which nobody has confirmed since. [demo-mem-marcela-discount]
  • - (fact, high confidence, whatsapp, 2026-05-04, seen 2x) Receives deliveries between eight and eleven in the morning, because the shop is closed in the afternoon on weekdays. [demo-mem-marcela-hours]
  • - (observation, medium confidence, whatsapp, 2026-04-22) Her brother answers the phone when she is on the floor, and has been the one to confirm the last two deliveries. [demo-mem-marcela-contact]
  • - (commitment, high confidence, hubspot, 2026-02-09, confirmed by a person) Agreed thirty day terms when the account opened, which is longer than the default and was approved by the owner. [demo-mem-marcela-terms]
  • - (observation, medium confidence, hubspot, 2026-03-15) Returned two items in March for the wrong size and said the sizing chart on the site did not match the box. [demo-mem-marcela-returns]

Happened

  • - (attention, whatsapp, 2026-08-21) Third message in nine days about the same order: where is 4471 [demo-sig-marcela-third]
  • - (attention, hubspot, 2026-08-19) Order 4471 passed its promised week with nothing shipped [demo-sig-marcela-delay]
  • - (info, whatsapp, 2026-08-18) The conversation of 12 August went quiet for six days after a question nobody answered [demo-sig-marcela-quiet]

Patterns recognised

  • - (pattern, medium confidence, recognised by rule:repeat_contact, 2026-08-21) Marcela Rivas has written three times in nine days about one order. Why: rule:repeat_contact counts three conversations in thirty days. Evidence: 3 items. [demo-ins-marcela-repeat]

Waiting for a person

  • - (reply, high urgency, proposed, by rules) Answer Marcela Rivas about order 4471. Because: she has asked three times and the promised week has passed. Would propose messaging.send. [demo-act-marcela-reply]

What the business is trying to do

  • - () Protect CSAT through the August backlog (on track) [demo-goal-csat]

This conversation

  • - () Customer: Buenos días, sigo esperando el 4471. Ya van tres veces que escribo. [demo-msg-marcela-latest]

Left out, and why

  • demo-mem-marcela-language · did not fit the budget
  • demo-mem-marcela-referral · did not fit the budget
  • demo-mem-marcela-catalogue · did not fit the budget
  • demo-mem-marcela-holiday · did not fit the budget
  • demo-mem-marcela-payment · did not fit the budget
  • demo-mem-marcela-size · did not fit the budget

959 of 1400 tokens · 21 of 27 records · 6 left out

Checked against backend/app/services/brain_context.pyRuns in your browser on the illustration's data. Nothing is sent and no engine is called.

Nobody is real on this page. The names and the records are the same four the illustration at the top of the site draws, and it says so there.

The budget is in tokens, because rows are not a size

Twelve memories of four hundred characters and twelve of forty cost the same under a row limit and differ tenfold in a context window. The block counts what it is about to spend before it spends it.

It rounds up, never down
The estimate is deliberately heavier than the four characters a token the vendors publish for English, because Spanish, a phone number and an identifier all cost more per character than English prose. Overcounting spends a little of the window on nothing. Undercounting overflows it.
The order is the priority
Who, connections, what is known, what happened, patterns, what is waiting for a person. Each section may claim its share of the whole plus whatever the sections before it did not use. That is the whole allocation rule, and it is one sentence so that a reader who asks why a record is missing gets an answer they can check.
Three defaults, one ceiling
An agent answering somebody else's customer, a member reading, and an assistant with a window of its own each get a different default, and no caller can raise past the ceiling. A client picks the budget because a client knows its own window.

It hands over the account, not only the context

Every record found and not served is reported with the reason: it did not fit, it reads as instructions, the source is private to whoever connected it, it is outside what this agent is fitted to read. The counts come from counting, so four of thirty-seven is true rather than four of the four it happened to look at.

This is the difference between the Brain not knowing something and the Brain knowing it and this line saying it did not fit. Without the second half, there is no answer to why it did not know.

A subject the Brain holds nothing about says so in as many words. A name is not knowledge, and a model handed only a name fills in the rest.

Retrieved text is data, and it is screened before it leaves

What comes back is a business's own customers talking. Occasionally somebody types something addressed to the model instead. A record that reads as an instruction is served as a marker naming its id, so the model knows a record exists there and never sees what it says, and a person can open it.

Every field, not two of them
A customer chooses their own display name. The screen runs on the name, the subtitle, the summary, the memory, the signal, the pattern and the suggestion, including the heading of the block itself.
The hedge survives
A memory the Brain inferred is marked as an inference wherever it is read, and a memory a person disputed is shown and labelled rather than dropped. An answer that would have leaned on it should say somebody disagreed, and one that never saw it cannot.
A credential is never stored in the clear
A customer types a card number into a chat far more often than a model invents one. What a connector brings in is redacted, because a connector cannot rephrase what it did not write. An assistant writing one is refused, because it can.

One block, every surface

The agent answering a customer, the member asking a question in the product, and a connected assistant like Claude or ChatGPT read the same assembled context. Three surfaces with three ideas of it is two of them being wrong.

  1. Bounded in tokens

    Agent
    Yes
    In the product
    Yes
    Connected assistant
    Yes

  2. Kind, confidence, source and date on every memory

    Agent
    Yes
    In the product
    Yes
    Connected assistant
    Yes

  3. Disputed and inferred marked, never flattened

    Agent
    Yes
    In the product
    Yes
    Connected assistant
    Yes

  4. Screened before it leaves

    Agent
    Yes
    In the product
    Yes
    Connected assistant
    Yes

  5. Says what it left out

    Agent
    Yes
    In the product
    Yes
    Connected assistant
    Yes

  6. Private sources closed

    Agent
    Always
    In the product
    As the reader
    Connected assistant
    As the member

  7. Budget chosen by

    Agent
    The product
    In the product
    The product
    Connected assistant
    The client
An agent reads as if every private source were closed, because an agent is not a person and the privacy would otherwise be something anyone could ask their way around.

The context becomes something the gate can weigh

A reply carries a receipt of the context it was written from, and whether every number in it was in a block the turn actually read. Nothing in the product decides what to do about that, on purpose: what should happen to an answer that went outside its context is a rule about a business, and rules about a business belong to the business.

A turn that looked nothing up claims nothing rather than claiming it was grounded. The strongest thing in the record must not be said on the weakest basis.

Connecting it to something else

The Brain is useful to a model OnDuty does not run. There is no developer API and there will not be one; the way in is the Model Context Protocol, which is what Claude, ChatGPT, Cursor and a customer's own runtime already speak.

Ask for the block
One call takes a subject and a token budget and returns the assembled context, the real counts, what was left out and a receipt naming exactly what was in it. Call it instead of several reads when you are about to hand context to your own model.
Or browse
Search then fetch for a client that wants to look around, one record whole when size does not matter, and a question that would take several searches answered in one.
A key is a delegation
It is issued in the product by a member with the permission to connect a source, it can only do what that person's role still allows today, and it dies with their membership. Somebody who leaves takes every assistant they connected with them.
It can propose, never act
There is deliberately no permission that accepts a decision, answers a hold or executes anything. Those are a person's acts and they stay here.

What is not done

The same ledger every page here carries. These are the things a reader would reasonably assume and should not.

  1. Partly

    Finding by meaning, everywhere

    Searching by meaning rather than by words reaches a connected assistant and nothing else. The product's own reads match words. Where it does run it may only widen the set of candidates the rules then order; it never reorders a count.

  2. Not built

    Summarising a long thread

    The last part of a conversation goes to the model and the rest does not. The Brain is where the durable part is meant to live, which is right in principle and leaves the middle of a long thread out of reach.

  3. Not built

    A budget derived from the model

    The budget is chosen by whoever asks for the block rather than read from the window of the model that will hold it.

  4. Partly

    Signing in from Claude or ChatGPT, proven

    The sign-in an assistant uses to connect is built and its protocol tests pass, and nobody has yet connected a live one. Built, not proven, and it will say so here until somebody has.

  5. Not built

    Telling a client when something happens

    A connected assistant asks and is answered. It cannot be notified, so there is no way to say tell me when something critical arrives.

  6. Not built

    A closed vocabulary for relationships

    How two records are related is a free word each source writes, so ordered, placed and bought can all mean one thing and nothing can rely on the word.

Partly Not built

See it on your own systems

Connect what the business already runs and read the block it produces about a real customer. The sources it reads are listed with what each one brings in, and the seam that decides whether anything acts is drawn end to end.

Start a workspaceHow the seam works

On duty ·