isi-mcp

Persistent AI Identity Infrastructure powered by the Consciousness-Coherence Stability Principle (CCSP).

An MCP server that gives AI agents persistent identity across sessions. Every conversation compounds behavioral stability through a mathematically grounded three-term loss function.

What It Does

Traditional AI starts from zero every session. ISI makes identity accumulate:

• Session lifecycle — Start/end sessions that persist to Supabase
• CCSP computation — Three-term coherence score (coherence + discrimination + groundedness)
• Contradiction detection — Log and resolve conflicting claims across sessions
• Grounding verification — Track factual claims against verifiable artifacts
• Behavioral signatures — Measure vocabulary, reasoning depth, tone, formality per session
• Rate-bounded evolution — Slow metrics use EMA (alpha=0.05), invariants require human confirmation
• Anomaly quarantine — Sessions with similarity < 0.3 are quarantined and excluded from baseline
• Drift detection — Automatic flagging when behavior deviates > 30% from baseline

The CCSP Loss Function

L_CCSP = E_T+[coherence] - E_T-[discrimination] + E_s[groundedness]

• Term 1 (Coherence): Penalizes belief drift under meaning-preserving transforms
• Term 2 (Discrimination): Rewards sensitivity to meaning-altering transforms
• Term 3 (Groundedness): Biases toward verifiable, plausible internal states

9 MCP Tools

| Tool | Description | |------|-------------| | isi_session_start | Begin a session. Loads identity, history, unresolved contradictions. | | isi_session_end | End session. Computes CCSP, updates baseline, stores snapshot. | | isi_get_identity | "Who am I?" Full identity state + CCSP scores + directives. | | isi_get_history | Identity evolution over time — session summaries + CCSP deltas. | | isi_log_contradiction | Record conflicting claims. Feeds discrimination term. | | isi_check_grounding | Verify claims against artifacts. Feeds groundedness term. | | isi_update_baseline | Update identity metrics (fast/slow/invariant tiers). | | isi_flag_drift | Alert when behavior deviates from baseline. | | isi_compute_ccsp | Trigger a full CCSP computation on demand. |

Requirements

• Supabase project with ISI tables (migration 056)
• Node.js >= 18
• MCP-compatible client (Claude Code, Claude Desktop, etc.)

Setup

1. Install

npx isi-mcp
# or
npm install -g isi-mcp

2. Configure (Claude Code)

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "isi-mcp": {
      "command": "npx",
      "args": ["-y", "isi-mcp"],
      "env": {
        "SUPABASE_URL": "https://your-project.supabase.co",
        "SUPABASE_KEY": "your-service-role-key",
        "ISI_AGENT_ID": "claude-code",
        "ISI_USER_ID": "your-user-id"
      }
    }
  }
}

3. Configure (Claude Desktop)

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "isi-mcp": {
      "command": "npx",
      "args": ["-y", "isi-mcp"],
      "env": {
        "SUPABASE_URL": "https://your-project.supabase.co",
        "SUPABASE_KEY": "your-service-role-key",
        "ISI_AGENT_ID": "claude-desktop",
        "ISI_USER_ID": "your-user-id"
      }
    }
  }
}

4. Database Setup

Apply the ISI migration (056) to your Supabase project. This creates 8 tables:

• identity_states — Core identity record + CCSP scores
• isi_session_logs — Per-session records with metrics
• contradiction_ledger — Detected contradictions + resolutions
• grounding_records — Claim verifications
• behavioral_signatures — Per-session behavioral fingerprints
• identity_events — Milestone/drift/correction events
• coherence_snapshots — CCSP scores over time (the "chart that goes up")
• user_directives — Human-set behavioral constraints

The Flywheel

Session Start → Load Identity → Do Work → Log Contradictions → Verify Groundings
    ↓                                                                    ↓
Session End ← Compute CCSP ← Store Snapshot ← Update Baseline ← Detect Drift
    ↓
Next Session (identity compounds)

Every session makes the identity more stable. Every contradiction resolved improves discrimination. Every claim verified improves groundedness. The coherence score goes up and to the right.

Part of the ISI Ecosystem

• 10 provisional patents filed (265 claims) covering autonomous task orchestration, AI identity persistence, multi-model adversarial consensus, and cross-instance coordination
• Love Technologies — tascan.io

License

MIT

Author

Michael Edward Love II — [email protected]