Closed-Loop State-Conditioned Generative Control System

Documentation • Paper • Cookbook

---

What is Entelechy?

Entelechy is a Memory-Weighted Policy Modulation Controller (MWPMC).

Most agent memory systems treat memory as a static "history" database that an LLM must actively choose to query. Entelechy inverts this paradigm. It acts as an invisible orchestrating middleware, treating the LLM not as an autonomous actor, but as a stateless functional unit that must be conditioned at every execution step.

Memory is not persisted as runtime state; rather, identity and policy are re-derived at every interaction from a weighted episodic memory graph (the State Reconstruction Loop) and used to modulate the LLM's inference-time controls.

The Zero-Intervention Architecture

Entelechy removes the burden of "tool use" from the LLM. It guarantees compliance and continuity through the SVT Continuation Protocol (SVT-CP), entirely mediated by your application backend (the Orchestrator).

The SVT-CP Loop:

1. Feature Extraction (Pre-Flight): Before the LLM generates a single token, the Orchestrator calls Entelechy's /bootstrap endpoint. Entelechy parses the user's prompt against their memory graph to compute a dense State Vector.
2. Policy Synthesis: Entelechy synthesizes the Policy Vector (Vector P), calculating required reasoning depth, verbosity, and disposition (Skepticism, Literalism, Empathy) based on historical context.
3. Policy Control (Injection): The Orchestrator invisibly injects this synthesized policy directly into the LLM's System Prompt. The LLM boots up completely aligned with the user's history.
4. Outcome Writeback: When the LLM finishes streaming its response to the user, the Orchestrator intercepts the payload and fires a background /retain_async webhook to Entelechy, logging the interaction as an [EXPERIENCE] fact. The memory graph evolves without the LLM ever making a tool call.

---

Frictionless Onboarding: Implicit 1-to-1 Mapping

Users should not have to manually configure "Brains" or manage memory infrastructure. Entelechy is designed to be deeply integrated with modern Auth providers (like Clerk, NextAuth, or Supabase).

When a user signs up to your application, your auth webhook simply calls:

POST /v1/default/banks/{user_id}

Entelechy's create_bank operation is an idempotent get-or-create. The user's isolated memory graph is provisioned instantly and invisibly.

---

The Control Plane

Entelechy ships with a Next.js Control Plane featuring a modern, cybernetic Coral-to-Amber GUI.

The Control Plane visualizes the SVT-CP execution pipeline and includes a built-in Zero-Intervention Chat Simulator. You can use the simulator to test your Orchestrator routing, observe the invisible /bootstrap injections, and watch background /retain_async operations evolve the graph in real-time.

---

Quick Start (Docker)

To launch the full Entelechy stack (PostgreSQL pgvector, Python API Engine, and Next.js Control Plane):

export ENTELECHY_API_LLM_PROVIDER=vertexai  # Or openai, anthropic, etc.
# Note: For Vertex AI, ensure your service-account-key.json is mounted or ADC is configured.

docker run --rm -it --pull always -p 8888:8888 -p 3005:3005 \
  -v $HOME/.entelechy-docker:/home/entelechy/.pg0 \
  ghcr.io/garybense/entelechy:latest

Dataplane API: http://localhost:8888
Control Plane UI: http://localhost:3005/banks/newdev?view=chat-simulator

---

Documentation & Developer Tools

🤖 Using a coding agent? Install the Entelechy documentation skill for instant access to the MWPMC architectural guidelines while you code:

npx skills add https://github.com/garybense/entelechy --skill entelechy-docs

Works with Claude Code, Cursor, Gemini, and other AI coding assistants.

Core Endpoints for Orchestrators

• POST /v1/default/banks/{bank_id}/sessions/bootstrap - Pre-flight State Vector extraction.
• POST /v1/default/memories/retain_async - Fire-and-forget background interception.
• GET /mcp/{bank_id}/ - Single-bank scoped MCP access for legacy agents.

License

MIT