Howdigitalcolleaguesremember

Two people in a meeting mention the third quarter decision, and both know what it means. They know the history, who was there and what was tried before. An artificial intelligence (AI) assistant dropped into the same conversation knows none of it. Context is what makes the difference between an assistant and a colleague, and most AI tools never get it.

Why most AI assistants forget

Most AI at work is attached to a document or a private chat window, and every conversation starts from nothing. The assistant knows nothing about the room, the project or the people, so the user has to explain it every time. That is fine for a one-off question and tiring for real work. It is one reason many pilots stop.

Digital colleagues work in a room that keeps the context. The members, the history and the linked knowledge belong to the room, and a colleague that joins receives all three. Nobody has to brief the colleague before each question, because the platform has already done it.

Chat has no sessions

A room is one long stream of messages with several topics running at once, and a follow-up question can be only four words long. So the platform keeps the memory, not the model. It puts every message into a topic. For each message it then prepares a short summary for the colleague: the most recent messages, a running summary, earlier messages that relate to the question, and the facts worth keeping.

Suppose you ask about the leave policy, get an answer, then ask "and for part-timers?". That question has no subject, but it still reaches whoever answered the first one. The room remembers every message from a person, including those nobody answered. A colleague added to a room today reads what happens from today, so it does not answer a question the team resolved last week.

Three levels of memory

Conversation memory is what the colleague needs for one thread. It is short-lived and detailed. Colleague memory is what one colleague has learned doing its job, such as decisions and preferences, so it improves over time instead of starting again. Team memory is shared by the colleagues and people working on a topic or project, so it is the organisation's memory.

Moving something from one level to another is always a deliberate step. Nothing goes from a conversation into team memory without a check in between. Checked knowledge works the same way: it is read-only, and changes go through a librarian.

Checked knowledge

People write documents for other people: narrative notes, scattered files and knowledge that only individuals remember. A colleague needs knowledge that is organised, can be cited and can be trusted. So knowledge goes through a checking step, from notes, repository documentation and processed documents into a library where every entry shows its source. The library is read-only. Changes go through a librarian colleague, either automatically, through feedback on a page, or by asking in chat. Every change has a recorded source and nothing is overwritten without a trace.

Our Atlas platform is the reference implementation of this approach, but you can use the approach without Atlas. You can use your own sources, start with a small set of checked documents, or use Atlas in full. A colleague uses the knowledge in the same way each time.

Open formats and standards

Proprietary document formats are designed for people to look at, which makes them hard for an agent that needs to read and write precisely. We use markdown and open diagram formats, so colleagues work on the content and people add the formatting at the end. For connections between systems we use open standards, the agent-to-agent protocol and the Model Context Protocol (MCP), instead of a proprietary agent interface. The room solves the harder problem, which is getting the right context in.