mem-save

# Memory Save Extract insights from the completed task and persist them to the knowledge system. ## Steps ### 1. Reflect on the Task Ask yourself: - What did I learn about the codebase that wasn't already in knowledge.md? - Did any hypothesis get confirmed or contradicted? - Did I discover a new pattern that should be tracked? - Did I make a mistake that revealed a rule worth recording? ### 2. Identify the Domain Match insights to existing domain folders in `knowledge/INDEX.md`. If no folder fits, create a new domain folder with all three files (knowledge.md, hypotheses.md, rules.md) and add it to INDEX.md. ### 3. Categorize and Write For each insight, decide where it goes: **knowledge.md** — Confirmed facts and patterns ``` - **[topic]**: [fact] (observed [date]) ``` **hypotheses.md** — Patterns that need more data ``` - **[topic]**: [hypothesis] (confirmations: N/5, first observed [date]) ``` **rules.md** — Confirmed 5+ times, apply by default ``` - **[topic]**: [rule] (promoted [date], confirmed N times) ``` ### 4. Promotion Check Scan `hypotheses.md` in relevant domains. If any hypothesis now has 5+ confirmations: - Remove it from hypotheses.md - Add it to rules.md with `(promoted [today's date], confirmed N times)` - Tell the user: "Promoted to rule: [hypothesis]" ### 5. Demotion Check If today's work contradicted an existing rule in rules.md: - Remove it from rules.md - Add it back to hypotheses.md with `(demoted [today's date], contradicted by: [reason])` - Reset confirmation count to 0 - Tell the user: "Demoted rule: [rule] — contradicted by [reason]" ### 6. Update Hypothesis Counts If today's work confirmed an existing hypothesis (but it's not yet at 5): - Increment its confirmation count - Tell the user: "Confirmed: [hypothesis] (now N/5)" ### 7. Summary Report what was saved: ``` Saved to [domain]: + N new knowledge entries + N new hypotheses ^ N hypothesis confirmations * N promotions to rules v N demotions from rules ``` ## Ask Before Saving Before writing, briefly list what you plan to save and ask: "Save these insights?" This gives the user a chance to add, remove, or correct entries before they're persisted. ## Writing Good Entries - Be specific and actionable — future sessions need to understand and apply these - Include dates so staleness can be assessed - One insight per bullet point - Don't duplicate existing entries — update them instead - Skip things derivable from reading the code — focus on non-obvious patterns and gotchas