lorekeeper-protocollisted

# Lorekeeper Protocol Follow this protocol every session to keep your memory store accurate, healthy, and growing. Inspired by MRAgent (Ji et al., NUS, ICML 2026): **memory is reconstructed, not retrieved.** Actively explore the memory graph — search, reason, follow links, prune dead ends, repeat. ## Phase 1 — Session Start **Trigger:** Beginning of every session, before any substantive work. 1. Identify the session topic (1–3 keywords). 2. Search: `lore_search({ query: "<topic>", min_score: 0.15, include_links: true })` 3. Read all returned memories — decisions, constraints, patterns. 4. Provide feedback: `lore_update({ memory_feedback: [{ id: "<id>", useful: true/false }] })` 5. If fewer than 3 results, run a broader fallback search. **Do not skip.** Working without context causes duplicate inserts and contradictory decisions. ## Phase 1.5 — Active Memory Reconstruction **Trigger:** After Phase 1, before acting on retrieved memories. Run the reconstruction loop until the surface is exhausted: ``` 1. READ top results. 2. REASON about what's still unknown. 3. TRAVERSE links from high-value memories. 4. SEARCH for inferred cues. 5. PRUNE dead ends. 6. REPEAT. BREAK when 2 consecutive iterations return no novel results. ``` **Rule:** If you've read a memory and still have open questions, run another search — don't settle for the first batch. ## Phase 2 — Mid-Session (Topic Shift) **Trigger:** Conversation shifts to a new domain, subsystem, or question. 1. Run a fr