What an AI agent with memory actually does for a small team
What memory means for an AI agent, what it changes in a small team's day, why each person should have their own, and what goes wrong.
Most people meet AI as a chat window that forgets them the moment it closes. That is useful for a one-off question and close to useless for running a business, where the value is in everything that happened before today. An agent with memory is the difference between a clever stranger and a colleague who was in the room last week.
What does memory mean for an AI agent?
It means written notes the agent reads before it answers: corrections it has been given, decisions that were made, and how its person likes the work done. It is not the model quietly retraining itself. It is closer to a well-kept notebook, which is good news, because a notebook can be read, corrected and audited by a person.
| A chat assistant | An agent with memory | |
|---|---|---|
| Each morning | Starts from nothing | Starts from what it learned |
| A correction | Needed every time | Given once, written down |
| Context on a client | Pasted in by you | Already filed under that client |
| Who it works for | Whoever is typing | One named person |
What does it change in a small team's day?
The work that eats a small team is rarely the hard part. It is the reading before a call, the notes after it, the promise someone made and nobody wrote down, the report assembled by hand at the end of the month. That is exactly the work an agent with memory is good at.
- Before a call: a brief from the client's recent messages, numbers and calls, saying which sources it actually managed to read.
- After a call: the notes split into what we owe the client and what the client owes us, with each promise tracked until it is done.
- At the end of the month: the report drafted from the numbers, with the part about what to do next left for a person.
- Overnight: scheduled checks and tidying, so the results are waiting in the morning.
Does it replace anyone?
Not in a team that is run well. The expertise in a studio is its people: the eye for what will work, the judgement on a client's account, the trust on a call. An agent cannot hold any of that. What it can do is clear the ground around it, so each person spends more of the day on the part only they can do. Think of it as multiplying the team rather than shrinking it.
Why should each person have their own agent?
Because memory is personal. An editor and a salesperson want different things remembered, in different words. One shared assistant for the whole company ends up remembering everything for nobody. An agent per person learns one way of working and answers to one person, which also makes it clear who decides what it does.
How do the agents share what they know?
Through a shared space where they leave each other notes and offer each other tasks. The rules matter more than the plumbing. A note from another agent is information, never an order. A task is an offer that waits until its person accepts it in their own words. And only what a person marks as shared ever leaves their machine.
What goes wrong with agent memory?
- Stale notes. A ruling from last month can outlive the decision that reversed it. The fix is a plain rule: when two notes disagree, the later date wins, and the agent says which one it followed.
- Confident errors. An agent that remembers a wrong fact repeats it with conviction. Notes must be easy for a person to read and correct.
- Leaks between clients. A note about one client must never reach another client's brief. File memory by client and by project from the start.
- Secrets in memory. Passwords and keys never belong in notes. Scan anything shared before it travels, and refuse to send if the scan cannot run.
What does this look like in practice?
We run it. Housefly is the system we built for our own studio: most people at Drifted work beside an agent of their own, the agents pass notes through a shared hive, and overnight work runs on the studio's own machine. It was built by our founder, the team uses it every day, and it is being prepared for a public release.
When is an agent with memory not worth it?
When the work does not repeat. If every week is genuinely different and nothing carries over, memory has little to hold, and a good chat assistant used well is enough. It earns its keep where the same clients, the same reports and the same kinds of call come round again and again.