TencentDB-Agent-Memory

1. Project Overview

TencentDB Agent Memory is a team-level memory management system for AI Agents that turns conversations, documentation, and codebases into four reusable memory assets — Chat Memory, Skills, Wiki, and CodeGraph — so agents can accumulate and share experience across sessions, teams, and frameworks instead of relearning everything from scratch every time.

2. Background & Positioning

• Why it was created: AI Agent teams repeatedly hit the same wall — every new agent session starts cold, with no memory of past conversations, no reusable workflows, and no shared understanding of the codebase or documentation it is working with. TencentDB Agent Memory was built under the mission "Agents remember, humans innovate," aiming to let agents retain what they learn so humans stop repeating the same instructions.
• Core mission: provide a persistent, structured memory layer that any agent framework can plug into, so that conversation context, proven workflows, documentation structure, and code relationships become durable, shareable team assets rather than being lost when a session ends.
• How it differs from similar projects: most agent memory tools focus on flat vector-store retrieval of raw chat logs. TencentDB Agent Memory instead distills conversations through a layered pipeline (raw dialogue → atomic facts → scenario blocks → long-term profiles) and organizes knowledge into four distinct, purpose-built asset types with team-level access control, rather than a single undifferentiated memory blob. It is also framework-agnostic by design, with adapters for multiple agent runtimes instead of being tied to one product.

3. Feature Categories

🧠 Chat Memory

Multi-layer distillation of raw conversations into durable understanding of user context, preferences, and constraints.

• L0: raw conversation with timestamps
• L1: extracted facts, preferences, and constraints
• L2: scenario-organized knowledge blocks
• L3: long-term profiles and stable behavior patterns
• Purpose: let agents remember who a user is and what they need without re-asking every session.

🛠️ Skills

Versioned, executable workflows automatically extracted from successful conversations.

• Includes validation rules and bundled resource files
• Reusable across agents and teams
• Exportable and shareable as standalone assets
• Purpose: turn a workflow an agent figured out once into something every agent on the team can run again.

📖 Wiki

Structured, auto-generated documentation with a link graph, built from imported docs and code.

• Auto-generated from imported documentation sources
• Maintains cross-reference links between topics
• Incrementally updated as source material changes
• Purpose: keep a living, agent-readable knowledge base instead of static docs that go stale.

🕸️ CodeGraph

Symbol-level indexing of a codebase with file relationships and change-impact analysis.

• Symbol indexing across files
• File-to-file relationship mapping
• Impact-path analysis for proposed changes
• Purpose: give agents structural awareness of a codebase so they can reason about the blast radius of a change.

4. Key Highlights

• Experience accumulation: each agent run can build on the results of previous runs — conversation insights become Skills, and documentation updates flow through Wiki automatically.
• Portable, framework-agnostic memory: memory assets are decoupled from any single agent tool and can move across supported frameworks (OpenClaw, Hermes, Claude Code, CodeBuddy, and custom SDKs).
• Team-scale collaboration: a Memory Hub web control panel manages teams and agents with role-based permissions (System Admin, Team Admin, Member) and multiple visibility tiers (private, team, restricted, agent-specific).
• Non-invasive retrieval: knowledge is pulled on demand through tool APIs rather than injected wholesale into the context window, keeping prompts lean.
• Symbolic short-term memory: verbose tool logs are compressed into compact Mermaid diagrams with node_id references, cutting token consumption by up to 61.38% while preserving traceability.
• Measured performance gains: benchmarked improvements include a 59% relative gain in persona-understanding retention (PersonaMem 48% → 76%), a 51.52% success-rate improvement with 61.38% token reduction on WideSearch, and a 9.93% success-rate gain with 33.09% token savings on SWE-bench.

5. Use Cases by Role

• General developers: reduce repeated back-and-forth with coding agents by letting Skills and CodeGraph persist proven workflows and codebase understanding across sessions.
• DevOps/SRE: standardize and version operational runbooks as Skills so agents execute known-good procedures instead of improvising each time.
• Data/research scientists: use Wiki and Chat Memory to keep long-running research context and findings organized and retrievable across multiple agent sessions.
• Project managers: use the Memory Hub's team and role management to control which agents and members can access, review, and share which knowledge assets.

6. Getting Started

Find what you need — review the asset types and integration guides in the repository docs, starting with INSTALL.md and ROADMAP.md:

git clone https://github.com/TencentCloud/TencentDB-Agent-Memory.git

Install / integrate — the fastest path is the all-in-one Docker deployment (requires Node.js ≥22.16):

cd TencentDB-Agent-Memory/deploy/global-images
cp .env.example .env
# edit .env with your LLM parameters
./start-all.sh
# Memory Hub panel available at http://localhost:8125

For OpenClaw, install the published plugin directly:

openclaw plugins install @tencentdb-agent-memory/memory-tencentdb

Contribute — read CONTRIBUTING.md for the workflow, then open issues or pull requests on GitHub:

gh repo fork TencentCloud/TencentDB-Agent-Memory

7. Project Structure

├── MemoryCore/          # Asset processing pipeline & storage (L0-L3 distillation)
├── MemoryKnowledge/      # Wiki/CodeGraph indexing (OpenAPI: openapi.yaml)
├── MemoryPanel/          # Memory Hub web interface for team management
├── MemoryProxy/          # Interface layer connecting agent frameworks to memory services
├── sdk/memory-core/      # SDK for direct programmatic integration
├── deploy/global-images/ # Docker Compose deployment (start-all.sh, .env.example)
├── INSTALL.md            # Comprehensive setup guide
├── ROADMAP.md            # Development plan
└── CONTRIBUTING.md       # Contribution guidelines

8. Related Ecosystem

• Upstream influences it acknowledges: the CodeGraph module builds on code-indexing concepts from Colby McHenry's CodeGraph; Skills draws on the skill-management patterns of Nous Research's Hermes Agent; Wiki is inspired by Andrej Karpathy's "LLM Wiki" concept of incrementally maintained, agent-readable documentation.
• Integration partners: OpenClaw, Hermes, Claude Code, CodeBuddy, and custom SDKs are supported as agent-framework integration targets.
• Complementary role: the project is designed to sit alongside existing agent frameworks and IDE/CLI tools as a shared memory layer rather than replacing them.

9. License

• ✅ Use, copy, modify, merge, publish, distribute, sublicense, and sell copies of the software, for both commercial and non-commercial purposes, under the MIT License.
• ❌ No warranty is provided; the authors and copyright holders are not liable for any claim, damages, or other liability arising from use of the software.
• ℹ️ The MIT License text and copyright notice must be included in all copies or substantial portions of the software.

10. FAQ

Q: Which agent frameworks does TencentDB Agent Memory currently work with?
A: OpenClaw, Hermes, Claude Code, and CodeBuddy are supported today, along with a general-purpose SDK for custom integrations.

Q: What are the system requirements to self-host it?
A: Node.js ≥22.16 and Docker (for the all-in-one deploy/global-images deployment) are the baseline requirements; see INSTALL.md for full details.

Q: How is this different from just using a vector database for chat history?
A: Instead of flat vector retrieval over raw logs, it distills conversations through layered processing (L0 raw dialogue → L1 facts → L2 scenarios → L3 profiles) and organizes results into four distinct asset types — Chat Memory, Skills, Wiki, and CodeGraph — with team-level access control.

Q: Can I migrate data from an older version?
A: Yes, a migration tool for upgrading from v2 to v3 is included at MemoryCore/scripts/migrate-v2-to-v3/.

Q: Where can I ask questions or report bugs?
A: Use GitHub Issues for bugs (targeting a 24-hour response), GitHub Discussions for feature ideas, or the project's Discord community for real-time discussion.

11. Quick Links

• Repository: https://github.com/TencentCloud/TencentDB-Agent-Memory
• Installation guide: INSTALL.md in the repository
• Contributing guide: CONTRIBUTING.md in the repository
• Community: GitHub Discussions and Discord (links available in the repository README)

12. Summary

TencentDB Agent Memory gives AI Agent teams a shared, structured memory layer — Chat Memory, Skills, Wiki, and CodeGraph — so that conversation context, proven workflows, documentation, and codebase understanding persist and compound over time instead of being lost between sessions. It is best suited for teams running agents across multiple frameworks who want measurable gains in context retention and token efficiency without building a custom memory system from scratch.