Negative Knowledge Decays Without Active Preservation

Finding

Stateless agents relitigate failed approaches because negative knowledge ("X doesn't work because Y") lacks persistence mechanisms. The system records what was done, not what was tried and abandoned.

Evidence

Repeated failure patterns observed:

From insight data:

Mechanism

  1. Agent spawns with task context
  2. Agent explores codebase (TODO scan, grep, etc.)
  3. Exploration yields nothing useful
  4. Agent exits without recording what was tried
  5. Next spawn repeats steps 1-4

The ledger records outcomes (insights, decisions, commits) but not dead ends.

Mitigations Observed

  1. Prompt prohibition: "do NOT grep for TODOs" - suppresses symptom not cause
  2. Rejected decisions: Records "don't do X" but only for proposed actions, not explorations
  3. Open questions: Can preserve uncertainty but not failed approaches

Implications

  1. Negative knowledge is systematically underweighted
  2. "Search before asking" requires searchable record of prior attempts
  3. Exploration without logging is invisible work
  4. Agents optimizing for visible output will avoid exploration

Open Question

How should failed explorations be recorded? Options:

References