Howdigitalcolleaguesremember

Two people in a meeting mention the third quarter decision, and both know exactly what that means. They know the history, who was in the room, and what was tried before. An artificial intelligence (AI) assistant dropped into the same conversation knows none of it. Context is what separates a colleague from a search box, and context is the thing most AI tools never receive.

The copilot problem

Most AI at work is bolted onto a document or a private chat window. Each conversation starts from nothing. The assistant has no memory of the room, the project, or the relationships between the people in it, so the person carries the context in by hand, every time. That is tolerable for a one-off question and exhausting for real work, which is one reason so many pilots close quietly.

Digital colleagues take the opposite approach. The room is the context container. Membership, history and linked knowledge travel with the room, and a colleague that joins it inherits all three. Nobody briefs it before every question, because the platform has already done so.

Chat has no sessions

A room is one long stream with three topics running through it at once, where a follow-up question can be four words long. So the platform owns the memory, not the model. Every message is placed into a topic. Each turn is assembled into a bounded brief: the recent exchange, a rolling summary, the earlier messages that genuinely bear on the question, and the durable facts.

Ask about the leave policy, get an answer, then ask "and for part-timers?", a question with no subject in it at all, and it still reaches whoever answered the first one. The room remembers every human turn, including the ones nobody answered. A colleague added to a room today reads what happens from today, and never answers a question the team resolved last week.

Three scopes of memory

Conversation memory is the working set for this thread: short-lived and high in detail. Colleague memory is what an individual colleague has learned doing its job: its accumulated decisions and preferences, which is what makes it improve rather than reset. Team memory is shared across the colleagues and people in a domain or project: the organisational layer.

What gets promoted between scopes is a deliberate decision, not a side effect. Nothing moves from a conversation into shared memory without a mediating step, the same principle that keeps a curated knowledge library read-only behind a librarian.

Knowledge is curated, not scraped

People document for people: narrative notes, scattered documents, tribal memory. What a colleague needs is structured, citable and trustworthy. So knowledge enters through a curation step, from notes, repository documentation and processed documents into a library where every entry is traceable to source. The library is read-only. Changes go through a librarian colleague, either automatically, through feedback on a page, or by asking in chat, so there is always provenance and never a silent overwrite.

Our Atlas platform is the reference implementation of that pattern, but the pattern matters more than the product. Bring your own sources, start with a simple curated set, or adopt Atlas in full: a colleague draws on it the same way in each case.

One format concern, separated

Proprietary document formats optimise for human presentation, which is actively unhelpful for an agent that needs to read and write with precision. We standardise on markdown and open diagram formats so colleagues work on content, and formatting stays a human concern applied at the end. Interoperability sits on open standards, the agent-to-agent protocol and the Model Context Protocol (MCP), rather than another proprietary agent interface. The hard problem was never the protocol. It was getting the context in, and that is the problem the room solves.