What's this about?
Diagnosis-guided self-evolution via hierarchical error memory — improve by first diagnosing what went wrong, and keep a structured memory of errors organised by level. 15 upvotes.
Why organise errors hierarchically?
Because "it failed" is not actionable and "line 47 threw a null pointer" does not generalise. A hierarchy separates the levels: a strategic error (wrong approach entirely), a tactical one (right approach, wrong sequence), an execution one (right step, wrong argument). Each demands a different correction, and lumping them together means the system learns the wrong lesson from the wrong level — patching a symptom when the plan was wrong, or rethinking the plan when it was a typo.
Why does this matter to you?
- "Diagnose before improving" is the same ordering argument as "evaluate before improving" from the rubric-induction paper.
- Error memory that persists across runs is the durable-value argument that WikiSkill and Warp's self-improving reviewer both make.
- The hierarchy is the contribution; flat error logs are what everyone already has and they do not help.