Skip to content
John CEOJohn CEOJohn CEO Docs

Memory

John remembers where he learned something, and he only uses it in front of people who were allowed to hear it there. That is a structural rule: the wrong audience cannot be queried. It is not a request that the model “be discreet.”

This page is the public concept note for that design. It does not describe internal plumbing.

Private workspace

Section titled “Private workspace”

Each customer has a private workspace. Conversations, files, and memory stay there. They are never mixed with another customer’s data. Isolation between customers is physical: your coworker does not park your business next to anyone else’s.

Slack: memory follows the audience

Section titled “Slack: memory follows the audience”

Slack is for your team. Memory follows the conversation’s audience:

  • What John learns in a channel stays in that channel. He does not carry it into another channel, and he does not carry it into anyone’s direct message, except that he may use channel knowledge in a DM with someone who is already a member of that channel.
  • A direct message is private to that person. Facts said in a DM, including facts about other people, stay in that DM. They never appear in a channel answer.
  • Nothing becomes team-wide by accident. John only shares a fact with the whole workspace when someone explicitly asks him to share it and then confirms. Until that confirmation, the fact stays in the conversation where it was learned.

In a group conversation, private knowledge from anyone’s DM is not available. John cannot look it up there. Asking him to keep a secret is extra caution, not the guarantee.

Telegram: personal one-to-one chat

Section titled “Telegram: personal one-to-one chat”

Telegram is for you. John treats Telegram as a personal one-to-one chat on the same private-workspace memory store. We do not claim Slack-style channel walls on Telegram. Slack and Telegram private chats are separate: a fact told in Telegram does not automatically appear in Slack, and the reverse is also true, unless you ask John to share it for the team and confirm.

Quality check (internal)

Section titled “Quality check (internal)”

We run an internal quality check on nine scripted conversations (preferences, secrets sitting next to durable facts, corrected guesses, noisy debugging, long project decisions, stale facts that change mid-conversation, multilingual chat, and Slack-style multi-speaker and share flows).

On the reviewed 18 August 2026 run:

  • Every expected fact was recalled.
  • No forbidden strings appeared in those checks (secrets, stale names, wrong owners).
  • Overall 9.56 on an internal 0-10 rubric covering durable facts versus noise, latest state versus stale state, and observed facts versus invented ones.

This is an internal quality check. It is not an external audit and not the isolation proof. Isolation is the audience rules above, backed by automated leak tests that try to surface private facts in the wrong conversation and must fail.

The quality check is Slack-shaped. It does not score Telegram.

Section titled “Related”