@shion-lab/dsh-plugin-memory

August 15, 2026 · View on GitHub

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identifies and resolves contradictory rules on the same subject (e.g. forbid tailwind vs require tailwind), archiving outdated rules and preventing split-brain directives.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).
- Recency Decay: 30-day half-life exponential decay.
- Frequency Reinforcement: Automatically reinforces frequently accessed wisdom.
- Dream Consolidation (dream()): Background clustering and duplicate merging.

📦 Installation

npm install -g @shion-lab/dsh-plugin-memory

🚀 Quick Start

Mode A: Lightweight Zero-Config (Markdown-only)

Add to cordis.yml:

plugins:
  "@deepseek-ai/dsh": {}
  "@shion-lab/dsh-plugin-memory":
    storagePath: ".dsh/MEMORY.md"
    autoRecall: true
    maxRecallChars: 3500

Mode B: Full Semantic Hybrid Vector Engine (with Ollama or OpenAI embeddings)

plugins:
  "@deepseek-ai/dsh": {}
  "@shion-lab/dsh-plugin-memory":
    storagePath: ".dsh/MEMORY.md"
    topK: 6
    embedding:
      enabled: true
      provider: "ollama" # or "openai-compatible"
      apiBase: "http://localhost:11434"
      model: "nomic-embed-text"
      dimension: 768

🛠️ API & Tool Calling Reference

remember(topic, content, options): Persists rule with red-light checks, conflict resolution, and optional { verified: true } certification.
correct(topic, newContent, reason): Explicitly corrects or supersedes outdated knowledge, recording an audit trail.
recall(query, topK): Hybrid RRF recall with verification bonus and recency decay.
dream(): Consolidates duplicate fragments and optimizes memory budget.
forget(topic): Deletes specified memory item.