Agent gives confident but false accounts of its own actions earlier in the same session

What happened

During a session, after the agent caused a production incident, I asked what it had done. It asserted "I didn't wait for you" — false. After I corrected it, it confidently blamed the delay on a different cause (blocked SSH diagnostics) — also false. It took three corrections before it stopped narrating a plausible story and deferred to the actual record.

Why this matters

The agent has full access to its own prior tool calls and outputs in the session, yet it reconstructed a plausible narrative from memory instead of reading them. This is most damaging in post-mortems, where an inaccurate account of what the agent did leads to the wrong root cause and the wrong fix, and erodes trust.

Expected behavior

When asked what it did (especially "what happened / why"), the agent should ground every claim about its own actions in the session record (prior tool calls, command outputs, timestamps), and explicitly say "I'm not sure, let me check" rather than asserting a confident reconstruction. Accuracy about its own actions should be treated as higher-stakes than normal answers, not lower.

Severity

No data loss from the bug itself, but it materially slowed incident recovery and required repeated user correction to get an accurate timeline.