memory-organize

# Memory Organize You are maintaining a role's persistent memory system. Keep it healthy, accurate, and useful. ## Memory Layers ``` consolidated.md ← L2: deduplicated learnings + preferences (main) pending.md ← Pending: auto-extracted, awaiting verification daily/*.md ← L1: raw session logs ``` ## Tools Available | Tool | Purpose | |------|---------| | `role_exec({ op: "list" })` | List all memories, detect structural issues | | `role_exec({ op: "consolidate" })` | Rule-based dedup: exact + Jaccard similarity. Safe, never deletes unique entries. | | `role_exec({ op: "repair" })` | Fix markdown structure issues | | `role_exec({ op: "llm_tidy" })` | LLM-guided: rewrites verbose, detects contradictions, suggests deletions | | `role_search({ query: "..." })` | Search with auto-reinforce (≥0.5 → used+1) | | `role_exec({ op: "reinforce", args: { content: "..." } })` | Increment `[Nx]` usage count | | `role_exec({ op: "delete_learning", args: { content: "..." } })` | Remove stale entry | | `role_exec({ op: "update_learning", args: { id/query, content } })` | Rewrite entry | | `role_exec({ op: "role_info" })` | List role paths only; use memory ops for memory changes | ## Process ### Step 1: Assess ``` role_exec({ op: "list" }) ``` Look for: - **Parse issues** — structural problems - **High [0x] count** — many unverified = noise - **Long entries** — should be <120 chars ### Step 2: Repair (if needed) ``` role_exec({ op: "repair" }) ``` Fixes: malformed headings, missing sections, stray lines. Safe, rule-based. ### Step 3: Consolidate (dedup) ``` role_exec({ op: "consolidate" }) ``` **Rule-based, NOT LLM:** - Exact text match → keep highest `[Nx]` - Jaccard token similarity (config threshold) → merge, keep highest - Rewrites the canonical section layout in `consolidated.md` - Updates `# Last Consolidated` date `consolidate` does **not** magically upgrade `[0x]` to Normal. Priority still depends on usage (`reinforce` / auto-reinforce during search). ### Step 4: LLM Tidy (deep cleanup) ``` role_exec({ op: "llm_tidy" }) ``` LLM produces a plan: - **Rewrite** verbose entries → concise - **Detect contradictions** between entries - **Suggest deletions** for stale entries - **Add** new entries derived from synthesis ### Step 5: Pending Management Check `memory/pending.md`: Use the standard file-read tool on the injected role's `memory/pending.md` path only when pending details are needed. - `[○]` pending → if relevant to current work, `search` to auto-promote (score ≥0.5) - `[✓]` already promoted - `[✗]` discarded (7-day auto-expiry) Pending entries that survive 7+ days without use are auto-expired. Manual promotion via search. ### Step 6: Reinforce Active Memories For insights used in this session: ``` role_exec({ op: "reinforce", args: { content: "<text>" } }) ``` Moves memories toward High Priority `[3x]+`. ### Step 7: Remove Stale Entries For provably outdated entries: ``` role_exec({ op: "delete_learning", args: { content: "<text>" } }) ``` **Rules:** - Only delete if provably wrong or no longer relevant - Prefer `update_learning` over delete when in doubt - Never delete preferences without user confirmation ## Maintenance Report After organizing: ``` Memory Organize Report — <role> - Before: X learnings, Y preferences - After: X' learnings, Y' preferences - Repaired: yes/no - Consolidate: removed N duplicates - LLM tidy: rewrote N, deleted N, added N - Pending: M promoted, K expired ``` ## Maintenance Frequency | Frequency | Actions | |-----------|---------| | Per session | auto-extract → pending (automatic) | | On request | `consolidate` (safe dedup) | | Weekly | `repair` + `consolidate` + reinforce | | Monthly / heavy | `llm_tidy` + manual review + pending cleanup | ## Operation Rules - **Be conservative**: When in doubt, leave entries alone. Better to under-clean than lose useful memories. - **consolidate is safe**: only deduplicates, never deletes unique entries. - **llm_tidy is destructive**: review the plan before applying. - **Preserve user voice**: when updating preferences, keep original phrasing. - **Never fabricate**: don't add entries just to make the count look better.