context.md

Context: Fibon Design Log — Memory System Study

Date: 2026-07-14 Goal: Read and analyze fibon design log chapters on memory systems for harness enhancement project

Sources

• Ch03: AI 怎麼「記得」你？ (Memory) — 36,574 chars
• Ch04: 為什麼 LLM 的自信，不能當成證據？ (Trust) — 37,591 chars
• Ch05: AI 可以修改自己的原始碼嗎？ (Self-Evolving) — 40,442 chars

Tool Used

• crawl4ai (HTTP strategy, no Playwright) — successfully fetched JS-rendered site

Key Insights for Hermes Harness Enhancement

1. Three card types: State, Event, Knowledge — our MEMORY.md is flat freeform
2. 3×6 multi-dimensional structure: 3 card types × 6 axes (time, scope, source, concept, confidence, relation)
3. State card rules: Supersede-Fork, Cardinality, Half-Life Decay, Conflict Arbitration
4. Five-way retrieval: vector + tags + fuzzy + temporal + graph
5. Key quote: "記憶系統真正嘅敵人唔係忘記，而係錯誤地一直記得"