Cold Memory Kernel

July 8, 2026 · View on GitHub

Hermes Agent Architecture Repository

What is this?

一个用于 AI Agent 的架构治理与记忆系统设计仓库。 An architecture repository for AI agent memory, governance, and capability design.

What problems does it solve? | What makes it different?

Most AI agent projects show you what they can do — how many models, tools, and features they support. This repo shows you how an agent should evolve — with evidence, governance, and discipline.

Problem	Solution
🔄 Agent 记忆写了但没用	Hot→Cold→Config→Runtime 四层架构
🎯 工具体积膨胀	Tool Gating + Register≠Expose 分离
⚡ Provider 选错/超时	Provider Health Score + 动态路由
📋 架构决策混乱	Hypothesis + Observation Week 证据驱动
🧩 能力扩展困难	Capability Registry + AP-001 提案

Who is it for?

正在开发 AI Agent / Coding Agent 的开发者。 Built for Hermes Agent.

Quick Demo

python demo.py
# → Shows current architecture status, hypotheses, and routing logic

Current Status

Phase:  Observation Week 1 (Feature Freeze)
Status: Collecting real usage data before architecture decisions

🏗️ Architecture at a Glance

  User Query
       │
       ▼
┌─────────────────────────────────────────────┐
│              Router                          │
│  DeepSeek → Kimi → Xiaomi → 小Q (fallback)  │
└──────┬──────────┬──────────┬────────────────┘
       ▼          ▼          ▼
┌──────────┐ ┌──────────┐ ┌──────┐
│ DeepSeek │ │   Kimi   │ │Xiaomi│
│ (Primary)│ │(Analyst) │ │(Long)│
│ tool acc │ │ no tools │ │  1M  │
│   ess    │ │  deep    │ │ ctx  │
└──────────┘ └──────────┘ └──────┘
     ┌──────┐
     │ 小Q  │
     │(Local│
     │  0$) │
     └──────┘

Memory Layers:
  🔥 Hot    (2,200 chars) → Route hints, recent corrections
  ❄️ Cold   (unlimited)   → Structured knowledge, verified facts
  ⚙️ Config (.env)        → API keys, executable truths
  ⏳ Runtime (session)    → Tool outputs, ephemeral state

🎯 Governance Model

Type	What it means	Example
Constraint 🔒	Must follow, no debate	Register ≠ Expose separation
Hypothesis 🔬	Awaiting evidence	H-004: Tool Gating, H-005: Provider Health Score
Backlog 📋	Queued for future	Skill Discovery, Hook System
Proposal 📐	Complete design	AP-001: Capability-Driven Architecture

Lifecycle: Idea → Proposal → Backlog → Hypothesis Review → Hypothesis → Validated ❌✅

🧪 Current Hypotheses (H-001 ~ H-005)

ID	Hypothesis	Status
H-001	Search Policy — when to search vs. use internal knowledge	`provisional`
H-002	Safety Policy — tiered safety response	`provisional`
H-003	Decision Principles — meta-principles for agent decisions	`provisional`
H-004	Tool Gating — expose tools by task capability, not default	`provisional`
H-005	Provider Health Score — dynamic routing by latency/success	`provisional`

Current Phase: 🔭 Observation Week (Feature Freeze) — collecting real usage data before any architecture decision.

📁 Repo Structure

├── README.md
├── ARCHITECTURE.md              ← Full system design
├── GOVERNANCE.md                ← 5-question governance framework
├── schema.yaml                  ← Memory schema definition
├── observation-checklist.md     ← Observation Week tracking
│
├── architecture/
│   ├── hypotheses/              ← H-001 through H-005
│   ├── proposals/               ← AP-001: Capability architecture
│   └── backlog/                 ← Future ideas & references
│
├── notes/
│   └── c-language/              ← C language study notes (Weng Kai MOOC)
│       └── ch1-2/               ← Chapters 1-2 review with Q&A
│
└── references/
    └── claude-code-architecture.md

💡 Design Principles

1. 只记不可推导的信息    →  Only remember what can't be re-derived
2. 每层抽象必须消除≥2个问题 →  Each abstraction must solve ≥2 real problems
3. 证据驱动，非功能驱动    →  Evidence-driven, not feature-driven
4. Capability优先         →  Capability first, adapter second, software last

📚 Also Includes: C Language Study Notes

Following Weng Kai's "C Programming" MOOC (Zhejiang University). Each chapter: Knowledge Points → Questions → Answers with Explanations. Wrong-answer-first layout.

Chapter	Topics	Status
Ch1-2	Program structure, variables, data types, operators	✅ Done
Ch3	Conditionals and loops	⏳ Coming soon

🚀 Quick Start

git clone https://github.com/jim-688/cold-memory-kernel.git
cd cold-memory-kernel

# Start with the governance model
cat GOVERNANCE.md

# Explore the hypotheses
cat architecture/hypotheses/H-004.json
cat architecture/hypotheses/H-005.json

# Read the full architecture
cat ARCHITECTURE.md

Built For


👤 User	Single-developer, student (Windows, Hermes Agent)
📐 Scale	Lightweight by design — principles general, implementation pragmatic
🛠️ Stack	Hermes Agent, DeepSeek/Kimi/Xiaomi/Ollama, VS Code/Git

📜 License

MIT

_{⭐ If this helped you think about agent architecture, consider starring the repo!}