Query past decisions, enforce guardrails, and track calibration - all in two API calls. v0.14.0: Multi-Agent Isolation, Live Deliberation, and FORGE Plugin.
Semantic search across past decisions. Find what worked, what failed, and why. Directional search by structure ("where did we use this pattern?") or function ("what solved this problem?").
Policy enforcement that prevents violations before they occur. Block high-stakes decisions with low confidence. Require code review for production changes.
Every query and guardrail check automatically linked to the resulting decision. Full provenance of how each choice was made - zero client changes needed.
Describe decisions by both structure (what it looks like) and function (what it solves). Inspired by Minsky's Society of Mind Ch 12 - connecting patterns to purposes.
Pre-decision query results automatically linked as lightweight graph edges. See which past decisions influenced each new choice, with semantic distance scores.
Track Brier scores, success rates, and confidence calibration over time. Reason-type analytics show which reasoning patterns predict success.
Your monitoring stack tracks latency, tokens, and errors. Cognition Engines tracks whether your agent's decisions are actually good.
14+ MCP tools for Claude Code, Claude Desktop, and OpenClaw. CSTP protocol over JSON-RPC 2.0 via HTTP. Framework-agnostic - works with any agent.
Every decision flows through this loop, creating a compounding record of organizational judgment.
| Phase | What happens |
|---|---|
| Fetch | Load context and past decisions |
| Orient | Check guardrails and constraints |
| Resolve | Decide and record with reasoning |
| Go | Execute |
| Extract | Evaluate outcomes and distill patterns |
Available as a Claude Code plugin with hooks, commands, and skills that automate the entire loop.
Two calls cover the full agent workflow:
Session start → get_session_context (cognitive context: profile, calibration, patterns)
Decision point → pre_action (query + guardrails + record in one call)The server auto-captures deliberation traces, bridge-definitions, and related decision links - zero client changes needed.
{
"allowed": true,
"similar_decisions": [...],
"guardrail_results": [...],
"calibration_context": {"brier_score": 0.024, "accuracy": 0.963},
"decision_id": "abc123"
}