TencentDB Agent Memory: Team Memory Hub for AI Agents
TencentDB Agent Memory is Tencent's self-hosted hub that turns agent conversations, docs, and code into four persistent assets — Chat Memory, Skill, Wiki, and CodeGraph. Reach for it if several agents share one codebase and need versioned, assignable context instead of a fresh explanation every session. Skip it if you want a lightweight RAG plugin: this needs three services running plus two LLM configs, and it's still labeled Beta.
Understanding TencentDB Agent Memory
TencentDB Agent Memory is a team-level memory hub for AI agents, built by Tencent in TypeScript. It distills conversations, documents, and codebases into four reusable asset types — Chat Memory, Skill, Wiki, and CodeGraph — each carrying ownership, versions, and access rules. It runs as three services (memory-core, memory-hub, proxy) with a web panel for assigning what each agent can see.
The Challenge of Repetitive Agent Work
Every new agent session starts blank: project context has to be re-explained, documents get re-read from page one, and a workflow that already worked has to be rediscovered. TencentDB Agent Memory targets that specific waste — not memory in the abstract, but the repeated cost of onboarding a fresh agent, or a fresh session of the same agent, onto work someone already did. Its own benchmark frames the gap: on PersonaMem, an agent went from 48% to 76% accuracy at applying prior user information once memory was turned on.
Core Memory Assets and Management
- ✓Four memory asset types: Chat Memory (preferences, facts, decisions), Skill (versioned, reusable procedures with trigger conditions and validation rules), Wiki (structured docs with a link graph), and CodeGraph (indexed code symbols, files, call relationships, and impact paths).
- ✓Layered memory model: raw conversations are saved as L0, then distilled via an async pipeline into L1 Atom (extracted facts), L2 Scenario (project-organized knowledge blocks), and L3 Core/Persona (long-term profiles).
- ✓Hybrid retrieval: L2/L3 give a fast context bootstrap by default; when specific facts are needed, BM25 keyword search plus vector retrieval combine via RRF and fall back to L1/L0, with results capped by item count, character budget, and timeout.
- ✓Fixed Binding + ACL access control: assets are private, team, restricted (User/Role/Agent ACL), or agent-scoped, so a Release Skill can go to just the Release Agent while an Architecture Wiki goes to every dev agent.
- ✓Cold-start import: existing codebases, docs, and past agent conversation sessions can be imported directly so a new agent team starts from accumulated experience instead of zero.
- ✓Two-tier roles: a global System Admin manages users, teams, and all asset types; team-level Admin and Member roles handle collaboration and access within a single team.
- ✓Agents discover and pull memory through an API — /v3/tools/list to see what's available, /v3/tools/call to read a Wiki page, source, or impact path — so docs and code stay outside the context window until actually needed.
Example: Building an Agent Team
- •A one-person company runs four agents with distinct roles — Scout (research), Builder (write code/build), Reviewer (test/find issues) — plus the human, each pulling only the memory assets relevant to its job instead of one shared prompt.
- •A Scout agent gets user-interview Chat Memory, a market-research Wiki, and a competitive-analysis Skill; a Builder agent gets a product Wiki, a project CodeGraph, and a feature-delivery Skill — different loadouts, less noise per agent.
- •A Reviewer agent is equipped with historical-incident Chat Memory plus a release-checklist Skill so release problems already solved once don't get re-discovered by whichever agent runs the next release.
- •Onboarding a new team member, or a new agent, means loading the team's existing assets rather than re-explaining the project from scratch.
Getting Started with Installation
Installation is documented: clone the repo, cd into deploy/global-images, copy .env.example to .env, fill in two sets of LLM parameters (one for the memory group, one for the proxy group), then run ./start-all.sh, which starts memory-core, memory-hub, and proxy together and prints a one-line command you can paste into Claude when it's done. The web panel opens at http://localhost:8125. Fuller docs — standalone Memory Hub deployment, Proxy plus Claude Code/CodeBuddy usage, stop/cleanup, and a port reference — live in INSTALL.md. If you're upgrading from v1.x or v0.x, a separate migration tool moves existing data to v2.0.0+; new installs can skip it.
Strengths
- ✓Four distinct asset types (Chat Memory, Skill, Wiki, CodeGraph) map to genuinely different data shapes instead of dumping everything into one chunked-text index.
- ✓Fine-grained sharing: private/team/restricted/agent visibility plus User/Role/Agent ACLs, not just a single shared-or-not toggle.
- ✓One-command bring-up: ./start-all.sh launches all three services and hands you a paste-ready command for Claude when it's done.
- ✓The PersonaMem benchmark (48% to 76%) at least gives one concrete, checkable number instead of just a claim of better memory.
- ✓Skills carry versions, trigger boundaries, and validation rules, not just free-text snippets, per the README.
Current Limitations and Roadmap
- △Labeled Beta in the README ("Team Memory Beta is evolving quickly") — features and APIs can still shift.
- △CodeGraph currently prioritizes public HTTPS repositories; private-repo and SSH credential support is still being refined, per the README's own notes.
- △Wiki and CodeGraph build asynchronously, so a newly imported doc set or codebase needs processing time before it reaches "ready" status — it isn't queryable instantly.
- △Memory routing across agents is still manual: the Hub supports binding assets by hand, but the README says fully automated routing is still under iteration.
- △Cross-framework support is currently limited to OpenClaw, Hermes, Claude Code, CodeBuddy, and SDK integration; broader framework coverage is on the roadmap, not shipped yet.
Comparing Agent Memory Approaches
Frequently Asked Questions
TencentDB Agent Memory is released under the MIT license, per the LICENSE file linked in the README.
TencentDB Agent Memory currently supports OpenClaw, Hermes, Claude Code, CodeBuddy, and direct SDK integration; the README lists broader cross-framework migration as a roadmap item, not yet shipped.
TencentDB Agent Memory ships a dedicated migration tool for moving data from older v1.x or v0.x releases to v2.0.0 and later; new installations don't need it.
TencentDB Agent Memory adds ownership, versioning, status, and team/ACL sharing on top of retrieval — standard RAG chunk-retrieves documents but doesn't track who can use a result, which version is current, or which agent should receive it, according to the project's own comparison table.
The README calls the current release "Team Memory Beta" and notes it's "evolving quickly," plus flags manual memory routing and limited private-repo CodeGraph support — treat it as pre-production unless you can absorb API and behavior changes.
Who should try it — and who should skip
Try TencentDB Agent Memory if you're running several agents, or agent frameworks, against a shared codebase or knowledge base and want assets — Skills, Wiki pages, CodeGraph — that persist and get assigned deliberately rather than re-injected into every prompt. Skip it if you need a single-agent, drop-in memory layer: standing up memory-core, memory-hub, and proxy, then configuring two sets of LLM parameters, is more infrastructure than a solo project usually needs. Teams already committed to OpenClaw, Hermes, Claude Code, or CodeBuddy get the most out of it today, since those are the only frameworks it names as supported.