Your model confabulates
If you ask someone with a gap in their memory what they did yesterday, they will often tell you, in detail and with complete confidence, about a day that never happened. They are not lying, the gap just fills itself with something plausible, and from the inside it feels exactly like remembering. Doctors call this confabulation. Your AI coding agent may do something very similar when it compacts.…
Your AI coding agent may fill in missing information about a day that never happened, a phenomenon known as confabulation in humans. Compaction replaces a session with a summary, turning live sources into remembered ones. If you ask a model what it did yesterday, it will often confidently recount an event that never occurred. The model may answer from memory instead of checking the source files.
Eight out of sixteen compacted sessions answered from memory, while both sessions without compacting went back to the file. The wrong answers arrived with zero or one tool calls, while the right ones had four or more. Compaction turns a live source into a remembered one, potentially leading to stale data. A controlled experiment was conducted with a set of configuration files, compacted sessions, and a change in the files' facts.
The model did not re-read the files, even when the facts in the files changed, indicating a potential mechanism of code staleness.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
