Endpoint Star Framework

A cognitive architecture for AI agents — not just a code library, but a framework for building agents that have memory, have judgment, and get better every time.

中文版 README | English

---

What is this?

A production-tested framework extracted from a real multi-agent system running daily since 2025. It provides four interlocking systems that together form a complete cognitive architecture:

System	Question it answers	What it does
Three-Doc System v2.0	"How to record"	Minimal-document project management with causal knowledge graphs
Thought Imprint v5.0	"How to judge"	Behavioral boundary principles in "first X, then Y" judgment-tendency format
Gene Module	"How to think"	Extracted cognitive instincts injected before skills as cognitive prerequisites
Promotion Chain	"How to evolve"	Self-improvement pipeline: lessons → skills → principles → constitution

Execute a task
    ↓ accumulate experience
lessons (experience)
    ↓ same-topic ≥3 times
skills (procedures)           ←── Gene module injects cognitive ability
    ↓ repeatedly validated + cross-domain
Thought Imprint (behavioral principles)
    ↓ stable ≥30 days + proven fundamental
SOUL Constitution

---

Quick Start

1. Read the cognitive architecture overview

Start with docs/cognitive-architecture.md — it explains how the four systems connect.

2. Pick what you need

If you want to...	Read
Structure your agent project documentation	docs/three-doc-system-v2.md
Define behavioral boundaries for your agent	docs/thought-imprint-v5.md
Extract cognitive patterns from agent operation	docs/gene-module.md
Build a self-improving knowledge pipeline	examples/promotion-chain-example.md
Set up multi-agent collaboration	docs/three-system.md

3. Copy templates

# Copy the agent templates
cp templates/AGENTS.md your-project/
cp templates/SOUL.md your-project/
cp templates/TOOLS.md your-project/

# Copy the gene module template
cp genes/GENES-template.md your-project/GENES.md

# Copy the system core template
cp examples/SYSTEM_CORE-template.md your-project/AGENT_SYSTEM_CORE.md

4. Fill in your details

• SOUL.md — Define your agent's identity, axioms, and constraints
• AGENTS.md — Define workspace rules and skill framework
• GENES.md — Start empty; genes are extracted after ≥30 days of operation
• AGENT_SYSTEM_CORE.md — Fill in architecture, leave "Cognitive Genes" as "Observation period"

---

Directory Structure

endpointstar-framework/
├── docs/                           # Methodology documents
│   ├── cognitive-architecture.md   # ★ Four-system unified specification
│   ├── three-doc-system-v2.md     # ★ Three-document system v2.0
│   ├── thought-imprint-v5.md      # ★ Thought imprint v5.0
│   ├── gene-module.md             # ★ Gene module application architecture
│   ├── three-system.md            # Three-system collaboration architecture
│   ├── multi-agent.md             # Multi-agent coordination
│   └── inter-brain.md             # Inter-brain communication protocol
├── templates/                      # Sanitized configuration templates
│   ├── AGENTS.md                  # Workspace specification
│   ├── SOUL.md                    # Personality definition
│   ├── TOOLS.md                   # Tool configuration (fill in your info)
│   ├── USER.md                    # User profile (fill in your info)
│   └── IDENTITY.md               # Agent identity
├── genes/                          # Cognitive gene module
│   └── GENES-template.md          # Gene template with examples
├── examples/                       # Worked examples
│   ├── SYSTEM_CORE-template.md    # System core with cognitive gene section
│   └── promotion-chain-example.md # How knowledge moves up the chain
├── lessons/                        # Production lessons learned
│   ├── README.md                  # Lesson index
│   ├── infrastructure.md          # Infrastructure (7 entries)
│   ├── ai-model.md                # AI model handling (8 entries)
│   ├── workflow.md                # Workflow patterns (10 entries)
│   └── security.md                # Security (8 entries)
└── CHANGELOG.md

---

Core Philosophy

The Cognitive Chain

Every agent action flows through a cognitive chain:

Thought Imprint → provides judgment tendency ("verify before concluding")
Gene Module → provides cognitive instinct ("unverifiable source → discard")
Three-Doc System → records execution process (writes to WORKFLOW.md)
Result:
   Success → may strengthen existing genes
   Failure → writes to lessons/ → may trigger promotion chain

Iron Rules

1. No fabrication — If you can't do it, say you can't
2. Reversible first — trash over rm, backup over direct edit
3. Infrastructure first — When search/channels/models break, stop and fix first
4. Write it down — "Mental notes" don't survive restarts; files do
5. Delegation closes the loop — Dispatched tasks must have timeout, feedback, fallback

The Memory Selection Standard

A piece of information enters long-term memory only if: without this record, the next conversation would not know about it.

• User preferences and decision logic → must record
• One-time debugging solutions → don't record
• Cross-project lessons and patterns → must record

---

Inspiration

This framework's philosophical foundation comes from Isaac Asimov's Foundation series:

• Endpoint Star (Terminus) = Knowledge preserver, seed of civilization rebuilding
• Seldon's Axiom = The agent's core behavioral principle
• Encyclopedia Galactica = The memory system
• Seldon Crisis = Every system crash and deployment failure

---

Origin

Extracted from a real multi-AI-agent production system running multiple agents (a primary dispatcher, a deep-reasoning agent, and a system-maintenance agent). The framework evolved through 3 iterations (v1.0 → v3.0) over 8 months of daily operation, driven by practical needs and hard-won lessons — including a critical "hallucination loop" failure where a fake r²=0.778 model was discovered and replaced with an honest r²=0.207 model.

---

License

MIT — Free to use, modify, and share.

---

Framework extracted from production AI Agent deployment | 2025-2026