@shion-lab/dsh-plugin-memory

Industrial-grade 4-tier persistent memory, self-correction, and cognitive defense engine for DeepSeek Harness (dsh).

Featuring Git-tracked Markdown storage, dense vector semantic search, red-light pollution guards, polarity conflict detection, and half-life recency decay.

Distilled from the battle-tested memory dynamics of desktop companion AI system Shion (紫苑).

---

🌟 Why dsh-plugin-memory? / 为什么需要立体持久记忆与自省防护？

By default, DeepSeek Harness operates in a stateless manner: once a session closes, all project-specific architectural rules, user preferences, and hard-earned debugging lessons are lost.

Furthermore, naive memory plugins suffer from memory pollution, hallucinations, and polarity conflicts (e.g. saving stack traces or conflicting rules).

@shion-lab/dsh-plugin-memory introduces a production-grade 4-tier cognitive architecture directly ported from companion agent practices (提炼自伴随式智能体「紫苑」实战记忆体系):

• 🛡️ Tier 1: MemoryGuard (Red-light Anti-Pollution Gate):
  • Strictly blocks stack traces (Traceback...), test runner output (pytest PASSED/FAILED), diff blocks, and API secrets (sk-...) from polluting memory.
• ⚖️ Tier 2: MemoryConflictDetector (Subject & Polarity Resolution):
  • Automatically identifies and resolves contradictory rules on the same subject (e.g. forbid tailwind vs require tailwind), archiving outdated rules and preventing split-brain directives.
• 📝 Tier 3: Git-Tracked Markdown Knowledge Base (.dsh/MEMORY.md):
  • Human-in-the-loop, version-controlled repository conventions with [✔ Verified] test-passed badges.
• 🧠 Tier 4: Dense Vector Hybrid Vitality Engine (.dsh/memory_store.json):
  • Dense Vectors: Cosine similarity semantic search (OpenAI / Ollama).