Goal Engine

English | 中文

Goal Engine is a persistent goal-management and behavioral evolution system for autonomous agents (specifically local OpenClaw agents). It transforms LLMs from "reactive predictors" into "goal-oriented evolutionaries" by enforcing an Attempt-Reflect-Evolve loop.

For a deeper dive into the vision and architecture, see: Introduction | 中文介绍

Core Philosophy

• Goal Persistence: Goals survive session restarts, failures, and environment resets.
• Continuous Evolution: Agents learn from failures via structured reflections and update internal policies.
• Autonomous Execution Stack: Provides mission compilation, strategy selection, boundary management, and no-false-done guards.

Detailed OpenClaw Integration

1. Prerequisites

• Node.js: >= 22.0.0
• pnpm: >= 9.0.0
• OpenClaw: A working local OpenClaw installation.

2. Installation

First, clone the repository and install dependencies:

git clone https://github.com/zhangdk/goal-engine.git
cd goal-engine
pnpm install

Start the Goal Engine service:

cd service
pnpm dev

Install the local OpenClaw plugin:

cd ..
./scripts/install-local.sh

This script registers the plugin, enables the necessary hooks (boot-md, bootstrap-extra-files), and configures the serviceUrl to http://localhost:3100.

3. Usage in OpenClaw

The Goal Engine exposes 5 core entrypoints to the OpenClaw agent. For best results, follow the Tool-First workflow.

A. Start a New Goal

When you give the agent a task (e.g., "Earn 100 RMB in a day"), it should compile it into a structured goal:

# Example invocation via adapter CLI
pnpm --dir agent-adapter openclaw entrypoint "start goal" --payload '{"title":"Earn 100 RMB in 24h","successCriteria":["Payment confirmation exists"],"currentStage":"initial"}'

B. Check Status & Alignment

Before any external action (search, browse, etc.), the agent must align the current task:

pnpm --dir agent-adapter openclaw entrypoint "show goal status" --payload '{"expectedGoalTitle":"Earn 100 RMB in 24h"}'

C. Record a Failed Attempt

If a strategy fails (e.g., blocked by a captcha), the agent must record it to update its policy:

pnpm --dir agent-adapter openclaw entrypoint "record failed attempt" --payload '{"stage":"initial","actionTaken":"Tried Baidu search","strategyTags":["search"],"failureType":"stuck_loop"}'

D. Check Retry (The Guard)

Before trying a similar path, the agent checks if the new plan is meaningfully different:

pnpm --dir agent-adapter openclaw entrypoint "check retry" --payload '{"plannedAction":"Try Bing search","whatChanged":"Switched from Baidu to Bing","strategyTags":["search"],"policyAcknowledged":true}'

E. Recover Goal Session

In a fresh session, use this to restore context from the service:

pnpm --dir agent-adapter openclaw entrypoint "recover current goal"

4. Example Workflow: "The Revenue Sprint"

1. User: "Give you a task: earn 100 RMB within one day, any method."
2. Agent: Calls goal_engine_start_goal to create the contract.
3. Agent: Selects a strategy (e.g., selling a micro-service).
4. Agent: Fails an attempt -> calls record_failed_attempt.
5. Goal Engine: Updates policy (e.g., "Avoid path X, try path Y").
6. Agent: Calls check_retry -> Allowed -> Executes new path.
7. Agent: Reaches success evidence -> Calls record_success (upcoming) or updates status.

Key Components

• service/: Node.js/Hono server backed by SQLite.
• agent-adapter/: Orchestration layer handling retry-guards and session recovery.
• openclaw/: Integration shell for the OpenClaw plugin ecosystem.

Access Local UI

Open http://localhost:3100/ui to observe agent evolution, evidence-based progress, and the learning timeline.

License

This project is licensed under the MIT License.