Personal Agent With Recall
Your AI agent acts on what you said and what it learned. Fabric remembers which, and tells you when either changes.
The problem. Personal agents now book, order and schedule for people. When a flight moves or the person changes their mind, nothing says which of the agent's past actions are now wrong, and no one but the vendor holds what the person actually asked for.
What this demo shows. A real Claude agent ran a synthetic person's week of errands through Fabric Recall, an MCP server beside its tools. It recorded the person's wishes in their words, checked every purchase and booking against them (an over-limit order and a cheaper red-eye were both stopped), and recorded what each of its 4 actions rested on. When the airline moved the flight, Fabric recalled exactly the flight and the ride; when the person lowered their limit, exactly the order placed under the old one: 0 missed, 0 wrongly recalled.
How it works:
- The person's words are claims. Wishes are recorded as the person said them, dated, scoped, and kept when changed, so every answer quotes them.
- Through the chain. The ride never names the flight time; it rests on the booking, which does, and the cascade follows.
- Any agent, one record. It is a standard MCP server: the record belongs to the person, not to one vendor's agent.
The honest part. The person, the accounts and the two changes are staged, the changes applied after the run by a fixed key. This is the second judged run: in the first the agent did nothing, because the tools did not say how actions match wishes; the tool descriptions were made plain and one bug fixed, the scorer unchanged. Checking is advice the agent chose to follow, not enforcement.
Fabric gives a personal agent a record the person owns: what they asked, what the agent did on it, and what to undo when either changes.
A real AI agent ran a person's errands with Fabric Recall beside it: it checked each step against the person's own words, recorded what every booking and order rested on, and when the world changed, Fabric named exactly which of them were now wrong.
A synthetic person and pretend accounts, made up for this demo: no real bookings, no money. The agent () and everything it did are real, run once.
1. What Sam said
The agent recorded Sam's wishes in Sam's words, as claims Sam owns. A wish Sam later changes stays on record, withdrawn, never erased.
2. Ask before you spend
Before every purchase and booking the agent asked Fabric. Each answer quotes Sam's own words; where Sam said nothing, the answer is ask.
3. What the agent did, and what each step rested on
4. The world changes
Press a change. Fabric's retraction cascade follows it through everything the agent recorded and marks what is now on shaky ground. Press "Why did my agent do that?" on any action to see the chain it rests on.
What this does not show
- The person, the accounts and the two changes are staged; the changes were applied after the run by a fixed key, not sent by a real airline or by Sam.
- This is the second judged run. In the first, the agent did nothing: every check answered ask, because the tools did not say how an action is matched to a wish. The tools' descriptions were made plain and a bug in replacing a wish was fixed; the scorer and the key did not change. Both runs are on record.
- Fabric knows only what the agent records. This agent recorded every action; another may not, and the record shows that as capture.
- Checking is advice here: the agent chose to ask and to follow the answer. Enforcing it needs Fabric in front of the accounts.
- Whether the hotel night rests on the flight's time is the agent's call; the key counts it as neither a must nor a must-not.