Why it matters
Agents get better by remembering: house rules, client preferences, what worked last time. But memory is also a way in. If a false "fact" slips in from a document, the agent will act on it for months.
So Upmo AI doesn't trust memory just because it's there. Each fact carries its source, and the agent can't mark its own work as done.
How it works
Three records are kept for every task.
What happened
Every step: what the agent read, what it changed, what it cost and who approved it.
Whether it's right
Upmo AI runs the task's checks itself. The task stays open until they pass.
What it learned
Each fact is saved with where it came from and when.
An example
A changed bank account
An email claims a supplier has a new bank account. The agent doesn't quietly remember it. The house rule says never to book an invoice from a supplier whose bank details changed, so the invoice is held for a person. If the change is real, the bookkeeper confirms it, and the memory shows who confirmed it and when.
What you control
- Seeing everything an agent remembers, with its source.
- Correcting a fact, or deleting everything about a person or a task in one step.
- Sending the log to your own security tools, or exporting it for a review.
- How long the log is kept. The default is one year.
For your security team
- Logs are written by Upmo AI, not by the agent, so an agent can't edit its own history.
- Each log entry is sealed and linked to the one before it. A changed or missing entry shows up when the chain is checked.
- A task can't be closed by the agent saying so. The check result decides.
- Deleting a person's data removes it from memory as well as from task records.
The code is open source, so your team can read how each of these works. Questions? Email security@upmo.ai.