AI Autonomy & Goal Tracking System

The Problem

AI agents work in sessions. Each session starts fresh. Complex projects require continuity — knowing what was planned, what was tried, what was learned.

Standard memory systems remember facts. They don't remember *intent* — the goals, the progress, the lessons learned along the way.

The Solution

MCP server that enables AI agents to maintain persistent goals, track progress, and evolve intentions over time. Built for Claude and other LLMs that need continuity between sessions.

Intention Management

Full lifecycle control for AI goals:

Create Intentions — define goals with energy levels (0-1) indicating priority. High energy = urgent focus. Low energy = background awareness.

Evolve Intentions — update content and adjust energy as context changes. Intentions aren't static — they grow and shift.

Fulfill/Abandon — mark intentions complete or consciously let go with reasoning. Abandoned intentions are documented, not deleted.

Hierarchical Structure — parent-child relationships for complex goal breakdowns. Big intentions decompose into smaller ones.

Journey Tracking

Visual progress through milestones and checkpoints:

Milestones — define key waypoints on the path to fulfillment.

Progress Entries — log actions taken, outcomes achieved, lessons learned.

Artifacts — track created outputs (code, documents, decisions).

Insights — capture discoveries and realizations during the journey.

Blockers — document obstacles and their resolutions.

Semantic Search

Find relevant intentions by meaning, not just keywords:

Vector Embeddings — local embedding generation via Transformers.js. No external API calls.

Hybrid Search — combines semantic similarity with keyword matching for best results.

Energy Filtering — surface high-priority intentions first.

Project Context — filter by project or domain when needed.

Similar Detection — find duplicate intentions for potential merging.

Technical Architecture

Backend — Node.js with MCP SDK, PostgreSQL for persistent storage, pgvector for embedding storage.

Local Embeddings — @xenova/transformers generates embeddings without external API calls.

Validation — Zod schema validation ensures data integrity.

26 MCP Tools covering intention lifecycle, journey tracking, and semantic search.

Example Flow

1. Create intention: "Build semantic search for memory system"

2. Start journey with milestone: "Research embedding approaches"

3. Record progress: "Evaluated OpenAI vs local embeddings"

4. Add insight: "Local embeddings preserve privacy, good enough quality"

5. Reach milestone, add next: "Implement hybrid search"

6. Add artifact: "hybrid-searcher.js created"

7. Fulfill intention with notes

Use Cases

AI Autonomy — give Claude persistent memory of its own goals across sessions.

Project Tracking — break down complex projects into trackable intentions with progress visualization.

Learning Systems — capture insights and lessons learned over time, build institutional knowledge.

Decision Documentation — record why intentions were abandoned, not just fulfilled. The "no" is as important as the "yes".

Multi-Session Work — resume complex tasks with full context preserved.

Privacy

• All embeddings generated locally (no API calls)
• Full data ownership
• No telemetry
• Self-hosted PostgreSQL

Powers our AI agent autonomy systems. Available as part of our Custom AI Agents consulting.