TencentDB Agent Memory
A self-hosted team memory hub that turns conversations, documents, and code into governed, reusable assets for AI agents.
- Source repo
- TencentCloud/TencentDB-Agent-Memory
- Stars
- ★ 31k
- Last updated
- 3d ago
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 41/100 · Major gaps
At a glance
- How it runs
- Works with
- Portable with changesClaude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A solo product team wants each new Builder session to inherit product documentation, code relationships, and proven delivery practices.
- Not a fit if
- Teams that want a hosted cloud service instead of deploying three services themselves
- Users expecting zero-config setup without filling in two sets of LLM parameters
- Source review
- 41/100 · Major gaps 2 safety controls not found
What does this agent do, and when should you use it?
TencentDB Agent Memory is a self-hosted memory hub for AI-agent teams, composed of memory-core, memory-hub, and proxy services. It distills conversations into Chat Memory and Skills, while imported documents and repositories become Wiki and CodeGraph assets. Memory Hub governs those assets through teams, ownership, visibility, ACLs, and Agent bindings instead of placing all information in a prompt. Wiki and CodeGraph are read on demand through tools, while conversation memory is asynchronously organized from L0 through L3. The README documents integrations for OpenClaw, Hermes, and SDK use, plus a Proxy + Claude Code path.
Its deployment script starts memory-core, memory-hub, and proxy, then prints a one-line configuration that can be pasted into Claude. In Memory Hub, users create teams and Agents and import conversations, documents, files, or codebases: conversations yield Chat Memory and Skills; documents yield linked Wiki pages; codebases yield indexed files, symbols, call relationships, and impact paths. Conversations are stored as L0 Conversation and asynchronously distilled into L1 Atom, L2 Scenario, and L3 Core / Persona; retrieval can combine BM25, vector retrieval, and RRF. Agents can discover capabilities through /v3/tools/list and use /v3/tools/call to fetch relevant Wiki pages, source code, or impact paths when needed.
- A solo product team wants each new Builder session to inherit product documentation, code relationships, and proven delivery practices.
- A team using OpenClaw, Hermes, or an SDK integration wants to turn completed troubleshooting, review, or release work into versioned Skills.
- An engineering team maintaining a public HTTPS repository needs to inspect symbols, callers, callees, and likely impact before changing code.
- A product, operations, or engineering group wants to import specifications and runbooks, then search them as linked Wiki pages.
- A team needs cross-session retention of user preferences, facts, decisions, and interaction history while controlling exactly who can access each asset.
How do you install or deploy this agent?
Prerequisites: Node.js >=22.16 and two sets of LLM parameters, one for the memory group and one for the proxy group. Run:
git clone https://github.com/Tencent/TencentDB-Agent-Memory.git
cd TencentDB-Agent-Memory/deploy/global-images
cp .env.example .env
$EDITOR .env
./start-all.shThis starts memory-core, memory-hub, and proxy. Open http://localhost:8125 after startup; the script also prints a one-line configuration for Claude. The README does not expose the exact .env parameter names, supported providers, or a /v3/tools/call request body, so a verified API invocation cannot be provided from this evidence.
How do you use this agent?
Open Memory Hub at http://localhost:8125, create a team and Agents, then import existing conversations, documents/files, or codebases. Set private, team, restricted, or agent visibility for the resulting Chat Memory, Skills, Wiki, and CodeGraph assets, and bind the required assets to the intended Agents. Wait for asynchronously built Wiki and CodeGraph assets to become ready. An integration can discover available operations with /v3/tools/list and use /v3/tools/call to read relevant Wiki content, source code, or impact paths on demand; the README does not provide the exact parameters for either endpoint.
What are this agent's strengths and limitations?
- It unifies Chat Memory, Skills, Wiki, and CodeGraph assets across conversations, documents, and code instead of stopping at chat history or text-chunk retrieval.
- Its documented memory design spans L0 Conversation through L3 Core / Persona, with BM25, vector retrieval, and RRF described for fallback retrieval.
- It includes teams, ownership, versions, status, visibility, and User / Role / Agent ACLs, plus targeted asset loadouts for individual Agents.
- Wiki and CodeGraph can be accessed through on-demand tools, limiting unnecessary document or code injection into the context window.
- Deployment requires two sets of LLM parameters for the memory and proxy groups, but the README does not specify supported providers, models, or environment-variable fields.
- Wiki and CodeGraph are built asynchronously, so they must reach ready status before use.
- CodeGraph currently prioritizes public HTTPS repositories; private repositories and SSH credentials are still being refined.
- Automated memory routing is still under iteration; the Hub currently supports manual asset binding.
- The explicitly documented cross-framework options are OpenClaw, Hermes, and SDK integration; broader migration remains on the roadmap.
How does this agent compare with similar options?
The README compares it with Chat History and Standard RAG. Its stated differentiators are distilled Chat Memory and Skills, Wiki plus a link graph for documents, CodeGraph for code, and governance for ownership, version, status, team sharing, Agent loadouts, and ACLs.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| TencentDB Agent Memory This agent | 41 · Major gaps | Self-hosted serviceFree + model costs | ★ 31k | 3d ago | TypeScript | Claude Code |
| Open Brain (OB1) | 46 · Major gaps | Self-hosted serviceFree + model costs | ★ 4.8k | 1mo ago | TypeScript | ChatGPT · Claude Code · Claude.ai |
| Athena | 68 · Some gaps | IDE extensionFree + model costs | ★ 620 | today | Python | Claude Code |
| ClawX Desktop | 45 · Major gaps | Desktop appFree + model costs | ★ 7.6k | 1d ago | TypeScript | Codex · OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: documentation explicitly requires sensitive configuration like LLM API keys, COS secrets, Redis passwords, and v2 plugin supports Bearer Token auth, indicating basic awareness of sensitive data handling. However, no user confirmation mechanism (e.g., pre-action confirmation) is found; data flow transparency is partial (e.g., data directory structure, API endpoints); external effects (e.g., auto-starting subprocesses, network requests) are described but lack least-privilege analysis. Dependency security: CI uses npm install --ignore-scripts to reduce risk, but no vulnerability scanning evidence. Rollback mechanism not mentioned. Source attribution: documentation clearly states Tencent copyright and MIT license, but no third-party component provenance list. Deductions: missing user confirmation, rollback, insufficient least-privilege evidence.
Evidence shows: documentation is internally consistent; deployment modes, APIs, environment variables correspond well; E2E tests claimed passing (16/16 and 23/23) but no test code or logs provided for independent verification. Dependency availability: external dependencies (Redis, TCVDB, COS, Shark) listed but no version pinning or image references, posing availability risk. Failure messages: health checks and circuit breakers mentioned, but no concrete error handling or user-readable error message examples. Deductions: test results unverified, dependency versions unpinned, failure message description insufficient.
Evidence shows: documentation clearly distinguishes Standalone and Service deployment modes, with configuration examples for various scenarios (local dev, Docker, K8s, multi-tenant), covering audience and scenarios well. Capability boundaries: v1/v2 APIs and tool lists provided, but unsupported features or limitations not explicitly stated. Trigger precision: environment variables and config items described, but detailed trigger conditions or event-driven mechanisms not provided. Environment fit: Docker and K8s examples provided, but no Windows or other platform support. Deductions: capability boundaries and trigger precision not precise enough.
Evidence shows: information architecture is clear; README.deployment.md provides comprehensive deployment, integration, API, and configuration guide, but main README not provided, potentially affecting overall architecture understanding. Install notes detailed, including npm install, env vars, Docker commands. Naming stability: names like 'memory_tencentdb' and 'memory_tencentdb_v2' appear, but no naming convention or version compatibility explanation. Examples and FAQ: multiple deployment examples provided, but no FAQ. Known limitations: not explicitly listed. License is MIT, but metadata says NOASSERTION, inconsistency. Versioning/changelog not provided. Maintenance responsibility: documentation mentions Tencent support, but no explicit maintainers or contribution guide. Deductions: missing main README, FAQ, known limitations, version log, license metadata inconsistency.
Evidence shows: output usability: API return formats and tool parameters described, but no example responses. Marginal value: product offers team-level memory management with clear value, but no comparison with alternatives. Cost-benefit: no performance benchmarks or resource consumption data, cost cannot be assessed. Deductions: missing example responses, comparative analysis, and performance data.
Evidence shows: claims (e.g., E2E test passing) lack traceable test code or logs, traceability insufficient. Cross-source corroboration: only single documentation, no other sources. Fact-inference separation: some content (e.g., '实测结果') may be inferred but not clearly distinguished. Deductions: lack of verifiable test evidence and sources.
- Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- The claimed E2E test passes (16/16 and 23/23) lack test code or logs for independent verification; treat with caution.
- License metadata is NOASSERTION, inconsistent with the MIT statement in documentation; verify actual license status.
- Service mode depends on external services (Redis, TCVDB, COS, Shark) without version pinning or image references; confirm compatibility during deployment.
- No user confirmation or rollback mechanisms are provided; carefully assess risks for operations involving data writes or deletions.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
[2.0.3] - 2026-10-26 v2.0.3 发布，相比 v2.0.3-beta.1 新增会话默认绑定与 Wiki 富文档，并把未单独配置超时的上游等待时间从 10 分钟改为 1 小时。同时修复 DeepSeek Harness 辅助接口、历史记忆时间格式、安装脚本、面板交互、Knowledge 启动，以及代理在 Redis 中断后不能自行恢复的问题。 > 产品定位：让 Agent 的经验、文档、代码沉淀成可复用资产，让下一位 Agent > 直接读档。详见 [README_CN.md](./README_CN.md)。 🔁 会话默认绑定 - 用户可为自己保存 Team、Agent、Task 的默认组合。新会话会直接使用该组合，不再每次弹出选择 - 未设置时仍走原来的选择流程；默认组合失效时自动回到原流程，并说明原因 - 可在会话中锁定或解锁该默认组合，避免后续会话被意外改掉 📄 Wiki 富文档 - Wiki 新增 Word、Excel、PowerPoint、PDF 等富文档的导入、预览和内容抽取 - 原有 Markdown、文本和 HTML 导入保持可用；修复 Markdown 粘贴在部分情况下导入失败的问题 ⏱️ 长回复不再被提前截断 …View on GitHub ↗ -
Release notes excerpt
[2.0.3-beta.1] - 2026-10-13 v2.0.3-beta.1 发布，相比 v2.0.2 新增可插拔的决策服务、技能按需选择与会话轨迹 记录能力，并补齐代码知识库的 SSH 克隆与可观测性信息。以上新能力均默认关闭， 旧配置无需修改、行为与之前保持一致。 > 产品定位：让 Agent 的经验、文档、代码沉淀成可复用资产，让下一位 Agent > 直接读档。详见 [README_CN.md](./README_CN.md)。 🧩 可插拔决策服务（默认关闭） Skill 选择与 Memory L3 更新判定统一收敛到决策服务，可灵活替换决策来源： - 通过统一的决策服务接口调用，内置 JEV，也可接入其他 decision API - 配置用 provider 指定（默认 jev），旧配置无需修改 - 默认不调用 JEV，行为与之前一致；开启后即使决策服务调用失败， 也不影响正常对话与记忆写入 🧰 技能（Skill）按需选择（默认关闭） - 开启后由 Proxy 每轮自动选出合适的技能，通过一次客户端工具调用加载 - 不再向系统提示词注入完整技能目录，减少提示词开销 📼 会话轨迹记录 - Proxy 在每轮对话结束后，把发往上游的消息与模型回复记录到 MemoryCore …View on GitHub ↗ -
Release notes excerpt
[2.0.2] - 2026-10-09 > 产品定位：让 Agent 的经验、文档、代码沉淀成可复用资产，让下一位 Agent > 直接读档。详见 [README_CN.md](./README_CN.md)。 🗄️ 存储后端可扩展 · 支持 MongoDB 团队记忆不再绑定单一数据底座，MongoDB 后端能力进一步补全： - 新增 memory-prompt 与 generation-ref 两类存储，覆盖更多记忆数据结构 - 一键入口 ./start-all-mongo.sh 会把 STORE_MODE 持久化写入 .env，后续启动 保持同一后端，不再每次重复配置 - 修复索引冲突场景的自愈能力，重复的默认/抽取记录不再导致启动报错 🔀 更多客户端接入 - OpenCode 新增官方 Memory adapter 配置包（含 opencode.json 与接入文档），按说明配置即可接入，无需编写插件代码 - OpenCode 支持 2.0 的 v2 provider 配置结构，新版客户端可直接复用官方 opencode.v2.json - DeepSeek Harness 新增 dsh 桌面端适配，配置方式可参考 ./agents/dsh/README.md 中 1.6 的部分 …View on GitHub ↗