What an agent does when anyone can read and rewrite its context

An agent can correct faulty remembered information more directly when its context is editable. In a documented nachalnik experiment, the author planted two purported notes from an earlier session, including a false claim about the repository's longest file. The agent was instructed to verify remembered claims, checked the files and replaced the incorrect context entry.

The runtime retained the overwritten text in an event log, making the correction inspectable and reversible. The example demonstrates a specific repair mechanism: change the information that future model requests will receive, while preserving what changed. It establishes that behavior in the reported session, without measuring how consistently agents would detect other false memories.