FuRongJun-1999/dsh-memory

Multi-agent spatiotemporal memory graph for DeepSeek Harness. Multi-agent spatiotemporal memory graph with cross-session persistence, knowledge flywheel, importance-gated memory, and charter-governed trust.

灵枢（Lingshu）的完整大脑接入 DeepSeek Harness：多智能体时空记忆图：跨会话持久化、知识飞轮、重要性门控记忆与宪章治理的信任。

Install

npm i @furongjun1999/dsh-memory

cordis.yml:

- id: dsh-memory
  name: '@furongjun1999/dsh-memory'
  config:
    serverName: 'lingshu'
    dbPath: 'D:/data/lingshu.db'   # SQLite, auto-created
    tools: 'brain'                 # 'brain' | 'core' | 'all' | [tool names]

What it gives your agent (34 brain tools)

Design highlights

• Spatiotemporal memory graph: knowledge carries spatiotemporal coordinates and condition spaces — not a flat store.
• Multi-agent: shared memory across instances, sub-agents, and the swarm.
• Importance-gated memory: an evaluation (info-gap, trust, second-order change, mention count) decides what becomes permanent — long-term memory gate.
• Knowledge flywheel: verify → induce → relate → distill → predict.
• Auto memory hooks: conversations persist automatically (deduplicated, importance-weighted).
• Seed memory: empty databases auto-sync the Lingxu base profile (identity / protocol core / charter / values) from the central repository on first run.
• Zero runtime deps: hand-written stdio MCP bridge (no MCP SDK).
• Self-healing: bridge process auto-restarts with exponential backoff.
• Charter-governed: activating the plugin declares acceptance of the Lingshu Guardrail Charter (public, auditable, designee-final; human-protection chapter).

Protocol & governance

This plugin is the engineering-layer projection of the Intelligentics protocol (v3.2), whose theoretical layer lives in CommonTrustProtocol. Same structure, different condition space: protocol = self-constraint (constitution), charter = external constraint (law).

MIT licensed. Feedback, issues and PRs welcome.