reasoning-bank

Confidence-weighted memory retrieval ranks memories using a confidence × recency decay formula. It searches project and agent memory directories with grep, parsing YAML frontmatter fields like confidence and last_used to surface the most relevant entries first.

Run a boost command to deliberately reinforce useful memories by updating their YAML frontmatter. The boost workflow deterministically increments the used_count, updates last_used, and applies a controlled diminishing-returns curve to the confidence field.

Flag low-confidence memories for pruning by keeping reinforcement and decay semantics aligned with your maintenance system. The workflow tracks older, low-confidence entries through frontmatter fields, producing ranked results that signal which entries are ready for eventual decay.

Yes, memory retrieval is guarded and does not auto-boost scores. The workflow explicitly separates retrieve, boost, and show actions, ensuring that searching for memories with grep does not deterministically update confidence, used_count, or last_used fields.

Memory retrieval works with missing YAML frontmatter fields by applying default values during parsing. It safely handles entries lacking confidence, used_count, or last_used fields, scoring them with defaults so the ranking results remain functional.

Flat memory retrieval returns noisy results when it lacks confidence × recency scoring. By parsing YAML frontmatter fields with defaults and applying a recency decay formula, the workflow ranks matches so the most relevant and reinforced memories surface first.