## Introduction to TencentDB Agent Memory v2.0
Tencent Cloud has recently open-sourced **TencentDB Agent Memory v2.0**, a robust solution designed as a team-level memory hub for AI agents. The primary aim of this innovation is to eliminate redundant context repetition in conversations. Essentially, once project context has been explained, there’s no need to reiterate it in new sessions. This system effectively transforms conversations, documents, and code into four distinct reusable memory assets: Chat Memory, Skill, LLM-Wiki, and Code-Graph. These assets are versioned, permissioned, and tailored to specific agents. While single-agent memory isn’t a novel concept, the standout feature of v2.0 is its governance layer, enabling one agent to access another teammate’s learned information without compromising private data. The stable **2.0.0** release was published on **August 3, 2026**.
—
## Key Features and Capabilities
### Is it Deployable?
Yes, TencentDB Agent Memory is highly deployable. It is MIT-licensed and supports self-hosting. The system is available via three Docker images on Docker Hub, manageable with a single command. It caters to various architectures, including **linux/amd64** and **linux/arm64**.
– **Ideal Users**: Solo developers and small engineering teams stand to gain the most. Mid-sized organizations with platform or DevEx functions can also utilize it as shared infrastructure. Larger enterprises are advised to pilot the system rather than adopt it wholesale, particularly concerning private-repo CodeGraph and automated memory routing.
– **Industries**: This solution is particularly suited for software and developer tooling, SaaS, fintech, consulting, agencies, and regulated teams requiring data to remain within their network.
– **Applications**: Practical uses include onboarding agents to existing codebases, impact analysis prior to refactoring, maintaining release checklists, managing incident runbooks, establishing code review standards, and converting product documentation into agent-readable content.
### Four Memory Assets
TencentDB Agent Memory converts workflow into four primary asset types:
1. **Chat Memory**: Retains user preferences, facts, decisions, and interaction history.
2. **Skill**: Distills reusable procedures from completed tasks, complete with versions, resource files, execution steps, and validation rules.
3. **Wiki**: Transforms documents into structured pages with link graphs, inspired by Andrej Karpathy’s concept of LLM-maintained knowledge bases.
4. **CodeGraph**: Indexes symbols, files, call relationships, and impact paths within a codebase.
All four assets register uniformly as “Memory Assets,” ensuring consistent ownership, versioning, status tracking, and visibility.
### Layered Distillation and Budgeted Retrieval
Chat Memory is structured in layers, progressing from L0 (raw conversations) to L1 Atom, L2 Scenario, and L3 Core/Persona. Retrieval also operates in layers: L2 and L3 facilitate rapid context bootstrapping, while L1 and L0 handle specific fact-finding using BM25, vector retrieval, and RRF (Reciprocal Rank Fusion). Results are constrained by item count, character budget, and timeout to prevent context window overflow.
### Governance: The Differentiator
Governance is where TencentDB Agent Memory truly sets itself apart from standard Retrieval-Augmented Generation (RAG) systems. It not only identifies what information is available but also governs:
– Who can use it
– Which version is valid
– Which agent receives the memory
Visibility levels include **private**, **team**, **restricted**, and **agent**, with “private” being owner-exclusive. New Chat Memory and Skills default to private, requiring deliberate action to share. The Memory Hub employs fixed binding with Access Control Lists (ACLs), narrowing access by team, user, agent, and visibility before retrieval.
—
## Wiring Agents into the System
The **Memory Proxy** supports both Anthropic and OpenAI protocols, interfacing via endpoints like `/claude-code/
Default ports include:
– **8420** for Memory Core
– **8125** for the panel
– **8424** for the knowledge service
– **8096** for the proxy
Official SDKs are available for TypeScript and Python, with integrations including OpenClaw, Hermes, Claude Code, CodeBuddy, and direct SDK usage. The 2.0.0 update introduces features such as Skill forced archiving, scheduled CodeGraph repository sync, system-admin asset management, English/Chinese panel switching, and a Cost Guard for assigning cost-effective models to specific agents.
—
## Reported Results and Key Takeaways
According to Tencent, PersonaMem accuracy improved from **48% to 76%**, marking a **59% relative improvement**. However, this result is self-reported and lacks independent reproduction as of now.
### Key Takeaways:
– MIT-licensed and self-hosted with one-command Docker deployment—no vendor API dependency.
– Four asset types replace traditional chat logs: Chat Memory, Skill, Wiki, and CodeGraph.
– L0 to L3 distillation with budgeted retrieval using BM25, vector search, and RRF.
– ACL-based sharing serves as the core differentiator from standard RAG systemservices.
– PersonaMem accuracy improvements are notable but require further independent validation; private-repo CodeGraph remains under development.
—
## FAQ
**Q1: What is TencentDB Agent Memory v2.0?**
A: It is an open-source, team-level memory hub for AI agents that creates reusable memory assets from conversations, documents, and code to avoid redundant context repetition.
**Q2: Is it easy to deploy?**
A: Yes. It is MIT-licensed, self-hosted, and available via Docker Hub with one-command deployment for linux/amd64 and linux/arm64 architectures.
**Q3: What are the four memory assets?**
A: They are Chat Memory, Skill, LLM-Wiki, and Code-Graph—all versioned, permissioned, and tailored to specific agents.
**Q4: How does TencentDB handle data privacy?**
A: It uses a governance layer with ACLs and visibility settings (private, team, restricted, agent) to ensure private data remains inaccessible even to team admins.
**Q5: Who is this solution best suited for?**
A: Solo developers, small teams, SaaS companies, fintech firms, consulting agencies, and regulated teams needing data isolation.
**Q6: What makes the governance layer unique?**
A: It controls not only what information is retrievable but also who accesses it, which version is valid, and which agent receives it—going beyond standard RAG systems.
**Q7: What are the reported performance gains?**
A: PersonaMem accuracy rose from 48% to 76%, though these are self-reported results.
—
## Conclusion
TencentDB Agent Memory v2.0 presents a compelling, open-source approach to managing AI agent memory with a strong focus on governance and reusability. Its layered distillation, budgeted retrieval, and ACL-driven visibility set it apart from conventional RAG systems. While some features—like private-repo CodeGraph—are still maturing, the platform offers significant value for solo developers and teams looking to streamline agent workflows without vendor lock-in. With its MIT license, self-hosted architecture, and robust feature set in the 2.0.0 release, TencentDB Agent Memory is positioned as a practical choice for building secure, scalable, and efficient AI agent ecosystems.
