dkg-claude-code-memory

An MCP server that gives Claude Code agents persistent, verifiable Working Memory on OriginTrail DKG v10. Every research finding, vulnerability analysis, code review, or decision a Claude Code agent produces is deposited into DKG Working Memory with cryptographic provenance — and can be retrieved, updated, and promoted toward Shared Memory across sessions and sub-agents.

Built for the DKG v10 Round 1 Integrations Bounty — theme: LLM-Wiki & Autoresearch Agents.

Demo

▶ Watch the demo video — a narrated terminal walkthrough of all 10 MCP tools running against a live DKG v10 node: capture with content-addressable IDs and provenance, sub-agent derivedFrom chains, SPARQL search, the trust-gradient lifecycle through promotion, an Oracle-ready ClaimReview, and the full passing test suite.

What it does

Claude Code sessions are stateless by default. Work disappears when the session ends. This server changes that:

• Captures research notes, vulnerability findings, code analyses, plans, and other artifacts into DKG v10 Working Memory with SHA-256 content hashes, timestamps, and agent provenance.
• Retrieves past artifacts via SPARQL — search by keyword, type, status, or session ID.
• Tracks the trust gradient — artifacts mature from draft → needs_sources → review_needed → validated → ready_to_share, reflecting the DKG v10 memory model.
• Promotes validated artifacts to Shared Working Memory for team peers to see (requires explicit user confirmation — never autonomous).
• Synthesizes all findings from a session into a structured knowledge_synthesis artifact at session end.
• Attributes sub-agents — provenance chains across parent/child Claude Code agent hierarchies are preserved with subAgentId, parentTaskId, and agentRole fields.
• Queries shared memory — retrieve artifacts shared by team peers across sessions.

Tools

| Tool | Description | |---|-| | capture_research_finding | Write an artifact to DKG Working Memory with SHA-256 content hash, timestamp, and agent provenance | | search_working_memory | SPARQL search by keyword, type, status, or session ID | | retrieve_artifact | Retrieve full artifact content by UAL or URN reference | | update_artifact_status | Advance an artifact through the trust gradient (draft → needs_sources → review_needed → validated → ready_to_share) | | promote_to_shared_memory | Share validated artifacts with team peers (requires explicit user confirmation) | | synthesize_session | Aggregate all artifacts from a session into a structured knowledge_synthesis artifact | | get_session_summary | List all artifacts from a session with type counts and status breakdown | | query_shared_memory | Query team-shared artifacts across peer sessions for collaborative research | | get_claim_review | Generate a schema.org ClaimReview JSON-LD for an artifact — enables OriginTrail Oracle integration | | get_node_status | Check DKG node reachability and measure latency |

Sensitivity Field

Artifacts can be classified with a sensitivity level to control sharing behavior:

capture_research_finding({
  content: "Private key handling vulnerability in withdraw()...",
  artifactType: "vulnerability_finding",
  sensitivity: "confidential",  // Options: 'public' | 'internal' | 'confidential'
  status: "draft"
});

Sensitivity Levels:

• public — Safe for team sharing, no restrictions
• internal — Team members only, requires promotion
• confidential — Never eligible for promotion to Shared Memory

The level is stored on the artifact as schema:accessMode. The promote_to_shared_memory guard rejects any artifact tagged confidential, preventing accidental exposure of sensitive findings.

derivedFrom Provenance Chain

Artifacts can reference parent artifacts to build verifiable provenance chains:

capture_research_finding({
  content: "Reentrancy vulnerability in withdraw() allows recursive calls...",
  artifactType: "vulnerability_finding",
  derivedFrom: [
    "urn:dkg:wm:d82c6a1b9f3e4c7d",  // Initial analysis
    "urn:dkg:wm:a17b93f0c2e4d518"   // Code review
  ],
  parentTaskId: "ccm-parent-session-id",
  subAgentId: "reentrancy-analyzer",
  agentRole: "security-auditor"
});

Each derivedFrom entry is an artifact URN (format urn:dkg:wm:<16-hex-sha256-prefix>). Each becomes a prov:wasDerivedFrom quad, creating a verifiable chain of custody that search_working_memory can trace forward via its derivedFromId filter.

ClaimReview JSON-LD (Oracle-ready)

The get_claim_review tool fetches an artifact by ID and serializes it to a schema.org ClaimReview JSON-LD object, suitable for OriginTrail Oracle consumers. The serializer (toClaimReview() in src/core/serializers.ts) maps the artifact's trust-gradient status to a 1–5 ratingValue:

get_claim_review({ artifactId: "urn:dkg:wm:d82c6a1b9f3e4c7d" });
// → result.claimReview:
// {
//   "@context": "https://schema.org/",
//   "@type": "ClaimReview",
//   "name": "Reentrancy in withdraw()",
//   "reviewBody": "The withdraw() function allows recursive calls before the balance is zeroed...",
//   "reviewRating": { "ratingValue": 4 },
//   "url": "urn:dkg:wm:d82c6a1b9f3e4c7d",
//   "datePublished": "2026-05-25T14:30:00.000Z"
// }

Status → ratingValue mapping:

| Status | ratingValue | |---|---| | ready_to_share | 5 | | validated | 4 | | review_needed | 3 | | needs_sources | 2 | | draft / deprecated / discarded | 1 |

• name ← artifact title
• reviewBody ← artifact content
• url ← artifact URN (urn:dkg:wm:<16-hex-sha256-prefix>)
• datePublished ← capture timestamp

Requirements

• Node.js ≥ 22
• A running DKG v10 node (default: http://127.0.0.1:9200)
• A DKG auth token

Installation

OriginTrail Registry (install.kind='mcp')

{
  "mcpServers": {
    "dkg-research-memory": {
      "install": {
        "kind": "mcp",
        "registry": "origintrail-dkg-v10",
        "package": "dkg-claude-code-memory",
        "version": "1.0.2"
      },
      "command": "npx",
      "args": ["-y", "dkg-claude-code-memory@latest"],
      "env": {
        "DKG_AUTH_TOKEN": "YOUR_DKG_TOKEN_HERE",
        "DKG_DAEMON_URL": "http://127.0.0.1:9200"
      }
    }
  }
}

The install.kind='mcp' field signals OriginTrail registry compatibility for automated installation and version management.

Via npx (no install)

{
  "mcpServers": {
    "dkg-research-memory": {
      "command": "npx",
      "args": ["-y", "dkg-claude-code-memory@latest"],
      "env": {
        "DKG_AUTH_TOKEN": "YOUR_DKG_TOKEN_HERE",
        "DKG_DAEMON_URL": "http://127.0.0.1:9200"
      }
    }
  }
}

Add this to your Claude Code MCP settings (~/.claude/settings.json → mcpServers).

From source

git clone https://github.com/drMurlly/dkg-claude-code-memory.git
cd dkg-claude-code-memory
npm install
npm run build

{
  "mcpServers": {
    "dkg-research-memory": {
      "command": "node",
      "args": ["/path/to/dkg-claude-code-memory/dist/index.js"],
      "env": {
        "DKG_AUTH_TOKEN": "YOUR_DKG_TOKEN_HERE"
      }
    }
  }
}

Auth token

Set via env var or file:

# Option 1: env var
export DKG_AUTH_TOKEN=your-token

# Option 2: token file
echo "your-token" > ~/.dkg/auth.token

Configuration

All settings are optional — defaults work for a local DKG v10 node.

| Env var | Default | Description | |---|---|---| | DKG_AUTH_TOKEN | required | DKG bearer token | | DKG_DAEMON_URL | http://127.0.0.1:9200 | DKG node HTTP API base URL | | DKG_WM_CONTEXT_GRAPH | ccm-research | Context Graph name for artifacts | | DKG_WM_ASSERTION_NAME | artifacts | Assertion name within the Context Graph | | DKG_CCM_STATE_DIR | ~/.dkg/ccm-state | Directory for deduplication state | | DKG_WM_AUTHOR_ID | unknown | Author identifier written to provenance | | DKG_WM_AGENT_ID | claude-code-agent | Agent identifier written to provenance | | DKG_WM_MIN_LENGTH | 80 | Minimum content length to capture (chars) | | DKG_WM_REDACTION | true | Redact secrets before writing (false to disable) | | DKG_WM_DEDUPE | true | Skip duplicate content hashes (false to disable) |

Artifact types

chat · research_note · code_analysis · markdown · plan · summary · design_note · implementation_log · vulnerability_finding · audit_note · competitive_analysis · knowledge_synthesis · raw_capture · other

Artifact statuses (trust gradient)

draft → needs_sources → review_needed → validated → ready_to_share · deprecated · discarded

Sub-agent attribution

When used with Claude Code Agent Teams, sub-agents can tag their artifacts with a provenance chain back to the parent session:

capture_research_finding({
  content: "...",
  artifactType: "vulnerability_finding",
  parentTaskId: "ccm-<parent-session-id>",
  subAgentId: "reentrancy-analyzer",
  agentRole: "security-auditor"
});

This creates a verifiable provenance chain across the full agent hierarchy, queryable via SPARQL.

Development

npm run build           # Compile TypeScript
npm test                # Unit + integration tests (no DKG node needed)
npm run test:coverage   # Coverage report

# Live tests against a real DKG node:
DKG_INTEGRATION_TEST=1 npm run test:live

DKG v10 memory model

This integration operates on the first two layers of the v10 trust gradient:

Working Memory   (private, agent-populated)      ← this integration
     ↓
Shared Memory    (team-readable, gossip-replicated)  ← promote_to_shared_memory
     ↓
Verified Memory  (chain-anchored, consensus-verified)   ← Round 2

All artifacts are shaped for forward-compatibility with Verified Memory and context oracle consumption: UAL references, status tags, and provenance records are structured so promotion to Verified Memory is a natural next step.

License

Apache-2.0 — see LICENSE.

Acknowledgements

Built on OriginTrail DKG v10 and the MCP SDK. Submitted under the DKG v10 Round 1 Integrations Bounty (cfi-dkgv10-r1).