Case study

An agent that sits in the company Slack and remembers what was agreed

We built our own agent for the daily running of the company. It sits in our Slack, keeps a long-term memory, watches reminders and answers from company documents. It has been running with us since the summer of 2026 and we are turning it into a product.

What it is
An agent in the company Slack, our own product
Status
Running with us since summer 2026
Operations
Run and watched by us
For customers
A dedicated instance, in build

What a company knows sits where nobody looks

A small company has no department keeping the overview. A decision is made in a thread, a note ends up in a document, a deadline in one person's head. A month later nobody can find any of it.

The assistants meant to solve this have two problems: they cannot see company data, and they forget after every conversation. Anyone who wants to use them has to paste the material in by hand - which is the very work they wanted rid of.

We wanted to try it on ourselves first. Without running it in our own company there is nothing to base a customer offer on.

An agent where the company already talks

The agent lives in Slack, because that is where we agree things anyway. There is no separate application to go to.

The memory is the interesting part. After each thread a cheap model judges whether anything was worth keeping and puts a note in a queue. Once an hour the queue is consolidated into topic files: rules, preferences, people, projects, decisions. The agent never touches direct messages.

  • Long-term memory with consolidation

    Capture after every thread and hourly consolidation into topic files. Every note records who prompted it.

  • Reminders without a model call

    The agent takes a deadline out of an ordinary sentence and brings it back in time. Delivery is watched by the system, not the model, so a reminder neither disappears nor arrives twice.

  • Answers from company documents

    The agent knows the uploaded documents and opens the right one when asked. No search machinery - for a small company's body of knowledge that is enough.

  • Learning through approval

    The agent does not write to its own long-term memory. It proposes a change and waits for a person. Approval means it knows; a rejection means nothing happened.

  • An admin console

    Knowledge, memory, reminders, tasks and the conversation archive in one place. What the agent knows can be read and deleted.

  • A fleet, not one assistant

    There are several agents and each has a role: incoming invoices, timesheets, operations. It follows the shape of the company rather than a list of features.

Where it stands today

every working day
in use in our Slack
within a minute
a reminder is delivered
only with approval
written to long-term memory

This describes our own operation, not customer results - the first customer instance is still being built. How much time it saves we do not yet measure.

Interested in an agent that knows your company?

Tell us what people in your company most often cannot find. We will tell you whether an agent is the answer or simpler software would do.

Talk about your problem