Claude wrote a wrong Pokémon landmark note and kept trusting it — Anthropic tested a memory repair pass

shoumikgoswami · reddit · 2026-11-08

In Anthropic's new memory write-up, Claude playing Pokémon wrote a landmark on the wrong side of a mountain in its notes, then kept acting on that mistake and backtracking.

The repair experiment:

An offline pass checked the note against the original session transcript and corrected it
The corrected run reached the next city; the uncorrected run never did

The author's extension: the useful thing to test in agents isn't whether they can save a fact, but what happens after they save a wrong one. For a coding agent, try planting an outdated setup instruction and checking whether later evidence actually replaces it rather than just adding a contradictory note.

Note this is an example from Anthropic's own testing, not a general reliability result.

Original post →

Claude wrote a wrong Pokémon landmark note and kept trusting it — Anthropic tested a memory repair pass

More from coding & agent