TencentDB-Agent-Memory: The Open-Source Memory Hub With 14,700+ GitHub Stars That Makes AI Agent Teams Remember Everything

What Is TencentDB-Agent-Memory?

Every developer who works with AI coding agents has experienced the frustration: you spend hours explaining your project architecture, coding conventions, and business logic to an agent — only to start from scratch in the next session. Multiply that across a team of agents, and the wasted effort becomes staggering.

TencentDB-Agent-Memory (also known as Memory Hub) is Tencent's open-source solution to this exact problem. It's a team-level memory infrastructure that automatically captures valuable knowledge from agent interactions and converts it into four types of reusable assets:

• Chat Memory — Preferences, decisions, facts, and interaction history distilled from conversations
• Skills — Proven workflows extracted from completed tasks, with versions and trigger rules
• LLM-Wiki — Structured documentation with link graphs (inspired by Karpathy's LLM knowledge base concept)
• Code-Graph — Indexed code symbols, files, call relationships, and impact analysis paths

The project has exploded on GitHub, earning 1,891 stars in a single day and accumulating over 14,700 total stars. The developer community's enthusiasm reflects a genuine pain point: AI agents are powerful, but they forget everything between sessions.

The Four Memory Asset Types Explained

1. Chat Memory — Cross-Session Understanding

Chat Memory goes far beyond simple conversation history. It uses a four-layer distillation pipeline to convert raw conversations into increasingly refined knowledge:

As the team explains it: "Don't refactor the old auth module — mobile is still using it." Context this costly shouldn't depend on humans repeating it every time. Chat Memory captures it automatically.

2. Skills — Distilled Executable Experience

Skills aren't just prompt snippets. They're full workflow definitions with:

• Version control and history
• Resource files and dependencies
• Trigger boundaries and activation conditions
• Step-by-step execution sequences
• Validation rules for quality assurance

Personal Skills are private by default. After review, they can be shared with the team and assigned to other agents. Think of it as: troubleshoot once, create a Skill, and the entire team benefits forever.

3. LLM-Wiki — Structured Knowledge Base

Wiki transforms product docs, design specs, and operational runbooks into structured pages connected by a link graph. Inspired by Andrej Karpathy's vision for LLM knowledge bases, it lets agents browse, search, and drill down through connected documentation — rather than re-reading entire file trees every session.

4. Code-Graph — Beyond "The Code Is Here"

CodeGraph indexes your codebase into a navigable graph of symbols, files, and call relationships. Agents can:

• Search for specific symbols and functions
• Inspect callers and callees
• Perform impact analysis before modifying code
• Understand which modules depend on each other

As the documentation states: CodeGraph doesn't just tell agents "the code is here" — it tells them "changing this might affect those."

How Memory Hub Compares to RAG

Many developers reach for RAG (Retrieval-Augmented Generation) when they need agent memory. But TencentDB-Agent-Memory goes significantly further:

The key insight: RAG answers "what can be found?" — Memory Hub also answers "who can use it, which version is valid, and which agent should receive it."

Real-World Example: Building a Team With Shared Memory

Imagine a small development team — let's call them Tiny but Serious Inc. — using multiple AI agents:

Tiny but Serious Inc.
├── 👤 You · Set goals / Make decisions
├── 🔭 Scout · Research / Find opportunities  
├── 🛠 Builder · Write code / Build products
├── 🧪 Reviewer · Test / Find issues
└── 🧠 Agent Memory · Preserve the team's experience

Each agent gets a tailored memory loadout:

🔭 Scout's Loadout:
 ├── User interview Chat Memory
 ├── Market research Wiki
 └── Competitive analysis Skill

🛠 Builder's Loadout:
 ├── Product Wiki  
 ├── Project CodeGraph
 └── Feature Delivery Skill

🧪 Reviewer's Loadout:
 ├── Historical incident Chat Memory
 ├── Project CodeGraph
 └── Release Checklist Skill

Different roles, different loadouts. Less noise — each agent gets the memory assets it actually needs. And when a new agent joins the team? It loads the team's "save file" on day one, inheriting months of accumulated experience instantly.

Getting Started in Minutes

Setting up TencentDB-Agent-Memory is straightforward. Clone the repo, configure your LLM parameters, and launch:

# Clone and configure
git clone https://github.com/Tencent/TencentDB-Agent-Memory.git
cd TencentDB-Agent-Memory/deploy/global-images
cp .env.example .env

# Edit .env with your LLM parameters (memory group + proxy group)
$EDITOR .env

# Launch everything with one command
./start-all.sh

Open the Memory Hub panel at http://localhost:8125 and start creating teams, importing codebases, and building your knowledge assets.

Access Control Model

The system uses a thoughtful visibility model:

Key Benefits

• Eliminates repetitive context-setting — Explain your project once, benefit forever
• Framework-agnostic portability — Memory assets work across Claude Code, Codex, and other coding agents
• Team knowledge compounding — Every session makes the next session better for all team members
• Cold-start friendly — New agents inherit existing experience instead of learning from scratch
• Granular access control — Private knowledge stays private; shared knowledge is explicitly shared
• Impact-aware code understanding — CodeGraph tells agents what might break before they make changes
• Self-improving over time — The async distillation pipeline continuously refines raw conversations into high-value assets

Why This Matters for the AI Agent Ecosystem

The AI agent landscape is shifting from single-agent workflows to multi-agent teams. Tools like Cloudflare Computer, LoopX, and others are giving agents more capabilities — but without shared memory, each agent operates in isolation.

TencentDB-Agent-Memory fills a critical gap: it's the connective tissue that lets agent teams accumulate experience rather than repeat it. As the project's philosophy states:

"Without Memory, loops may just repeat faster. With inherited memory, each iteration has the chance to be better than the last."

For development teams building with AI agents — whether using Claude Code, GitHub Copilot, Cursor, or custom frameworks — Memory Hub offers a way to turn the "learning cost" you've already paid into a permanent, shareable asset.

Ready to give your AI agent team a memory? Check out the project on GitHub and start building your team's knowledge base today.

Want to learn more about building with AI? Explore CoddyKit's courses to level up your development skills.

Frequently Asked Questions

What is TencentDB-Agent-Memory?

TencentDB-Agent-Memory (also called Memory Hub) is an open-source, team-level memory system for AI agents. It automatically extracts and organizes knowledge from agent conversations, documents, and code into four reusable asset types: Chat Memory, Skills, LLM-Wiki, and Code-Graph.

How is it different from RAG?

While RAG answers "what can be found?", Memory Hub also answers "who can use it, which version is valid, and which agent should receive it." It adds ownership tracking, version control, team sharing, and four specialized asset types beyond simple chunk-based retrieval.

Does it work with different AI coding agents?

Yes. Memory assets are framework-agnostic — they work with Claude Code, GitHub Copilot, Cursor, Codex, and custom agents. Switching frameworks only requires re-equipping the assets, not retraining.

How do I get started?

Clone the repo, configure LLM parameters in .env, and run ./start-all.sh. The Memory Hub panel opens at localhost:8125.

Is it suitable for solo developers?

Absolutely. Solo devs benefit from cross-session memory and Skills. Teams gain additional shared knowledge and access control. As the project says: "The company can be tiny. Experience can compound forever."