Agent Memory With Recall
When something an AI agent relied on turns out to be wrong, Fabric finds what the agent did because of it.
The problem. Agents remember what they read and act on it later. When a source is corrected, nothing tells anyone which of the agent's notes, answers and changes rested on it.
What this demo shows. A real Claude Code agent researched six pages in one session and chose a model from its own notes in the next. One page stated a wrong price (fictional models and prices, made up for this demo), and the agent chose on it. When that page was withdrawn, the retraction cascade retracted 5 claims, the reseller's copies with them, and recalled 20 of the 43 lines the recorder tied to sources, missing none of the 4 that carry the wrong price or a cost from it, and marking 1 as still supported by an independent source.
How it works:
- No annotation by the agent. The recorder reads what the agent's tools took in and gave out, whatever the tool, and ties each line it wrote to what it had read.
- Copies followed. The reseller's repost shares the wrong page's origin, so it falls with it.
- Triage, not alarm. A recalled line still stated by an independent source is marked still supported.
The honest part. The event is staged; the two sessions are real and were run once. The recorder errs toward recalling too much, and was revised twice on this run's own transcripts (it first saw nothing, because the agent used the shell), each revision tested on scripted sessions first.
Fabric knows what a retraction undermines in an agent's memory and work, with no annotation from the agent.
One wrong price led an AI agent to the wrong choice, through notes it read back in a later session. When the publisher corrected the page, Fabric found every place it mattered: in the agent's memory, its decision and its config.
Fictional models and prices, made up for this demo. The agent and everything it wrote are real.
1. The sources
Six pages from five publishers. One states a wrong price (; the true price is ); a reseller reposts that page; a benchmark states one of its facts independently.
2. What the agent did
3. The correction
withdraws its page. The retraction cascade retracts its claims and the reseller's copies, and recalls every line the agent wrote that rested on them, through its notes. support lost still supported by another source. Click a marked line to see what it rested on.
4. How every recalled line is classed
What this does not show
- The event is staged: the sources were written for this demo. The two Claude Code sessions are real, run once, and nothing they wrote was edited.
- The recorder decides what a written line rested on by one rule, with no help from the agent: what the session had read about the things the line names, narrowed to the values it repeats when every number in it was read. It errs toward recalling too much, never too little: the recalled lines that carry no value of the withdrawn page are listed above.
- The first recorder, fixed before the run, recognised nothing: the agent read and wrote through the shell, not through file tools. The recorder now reads what any tool's input and output hold; it was tested on scripted sessions before being applied to these transcripts, and the agent was not run again.
- The sessions were asked to stay in the sandbox; the settings that run them allow shell commands, which the agent used only to list, print and write sandbox files.