v0.15.0·Apache-2.0

Your agent makes decisions.It just doesn't remember them.

Cognition Engines gives AI agents a durable memory of their decisions — what they decided, why, and whether it worked. Agents consult precedent before acting, guardrails block known mistakes before they run, and calibration tracking shows whether the agent is genuinely improving or merely confident.

See the Decision Loop Read the Source

17: MCP tools; Model Context Protocol
34: CSTP methods; JSON-RPC over HTTP
1,400: Tests; 55 test files

The problem

Context windows forget. Consequences don't.

An agent that reasons well for one session and starts from zero in the next isn't accumulating judgement — it's re-deriving it, at cost, with a fresh chance of getting it wrong. The reasoning that mattered lives in a transcript nobody queries, and the outcome that would have graded it is never written down at all.

The gap isn't recall of facts. It's recall of judgement: the choice that was made, the alternatives that were rejected, the confidence attached to it, and what actually happened next.

What gets stored per decision

Decisionwhat was chosen
Contextthe situation and constraints in force
Reasonstyped and individually weighted
Stakes & confidencedeclared before the outcome is known
Guardrail resultwhat was checked, what it returned
Deliberation tracethe reasoning chain, captured live
Outcomelogged later — the grading signal

The loop

Five calls turn a stateless agent into one that carries judgement forward.

01
get_session_context
Session start
Agent profile, prior decisions, active guardrails, calibration.
02
pre_action
Decision point
Precedent, guardrails, calibration and patterns in one call.
03
record_thought
During work
Reasoning attaches to the open decision in real time.
04
update_decision
On completion
Finalises the record once the work has settled.
05
review_outcome
Later
Logs what happened — the entry that makes calibration possible.

Capabilities

Six subsystems, one record.

Decision memory

Retrieval

Decisions stored with context, reasons, stakes and confidence — retrieved by meaning or by exact wording.

Guardrails

Pre-execution policy

Declarative policies evaluated before an action runs, not after it fails.

Calibration

Brier scoring

Brier-scored confidence per category. Not “sounds sure” — has earned the right to be sure.

Deliberation traces

Chain-of-thought capture

The reasoning chain captured live and attached to the decision, not reconstructed from logs.

Decision graph

Typed relationships

Typed links between decisions that depend on, supersede or contradict each other.

Provenance & evidence

Audit export

Evidence mapped to control frameworks and exported as a verifiable bundle.

Full surface — 17 MCP tools and 34 CSTP methods →

Integration

Framework-agnostic by design.

Cognition Engines speaks MCP and JSON-RPC over HTTP. Anything that can make a request can use it — Claude Code, Claude Desktop, an agent framework, or raw curl. Storage is WAL-mode SQLite with FTS5; vector search runs on 2 pluggable backends behind a single abstraction.

Quick start →

.mcp.json

{
  "mcpServers": {
    "decisions": {
      "command": "npx",
      "args": [
        "mcp-remote@latest",
        "http://YOUR_HOST:8100/mcp",
        "--allow-http",
        "--header",
        "Authorization: Bearer YOUR_CSTP_TOKEN"
      ]
    }
  }
}

Give your agents a track record.

Self-hosted, Apache 2.0, no vendor account required. Clone it, run the demo stack, and point your agent at it.

Get Started on GitHub Governance & Evidence