TencentDB Agent Memory v2: own your team's agent context
Tencent open-sourced a team-level memory hub for AI coding agents under MIT. Self-hosted, private by default — and no workflow for when the memory is wrong.
Every developer on your team is re-explaining the same codebase to the same class of AI agent, several times a day. Tencent Cloud's answer is TencentDB Agent Memory v2.0 — an MIT-licensed, self-hosted memory hub that lets one engineer's agent read what another engineer's agent already figured out. It's the most interesting piece of agent infrastructure we've looked at this month, and the interesting part isn't the retrieval.
What actually happened
Tencent Cloud shipped the stable 2.0.0 release of TencentDB Agent Memory on August 3. It turns a team's working output into four distinct memory assets rather than one undifferentiated vector blob:
- Chat Memory — preferences, decisions and facts carried across sessions
- Skill — reusable workflows extracted from conversations, with versions and validation rules
- Wiki — structured doc pages with a link graph
- CodeGraph — indexed symbols, files, call relationships and impact-analysis paths
It's MIT-licensed and runs locally: clone, set your LLM parameters in .env, run ./start-all.sh, and memory-core, memory-hub and the proxy come up together with a panel on localhost:8125. Multi-arch Docker images cover amd64 and arm64.
The governance model is the part worth copying. Assets carry one of three visibility levels — private (owner only), team (visible to members), or restricted (explicit ACL). New Chat Memory and Skills are private by default; sharing is an action, not a leak. Team admins can't read private assets.
VentureBeat's coverage names the gap plainly: there's no established workflow for what happens when a shared memory is simply wrong. Take that seriously.
Why team agent memory matters for your business
The context is the asset, not the model. You will swap models three times in the next two years. The accumulated knowledge of why your inventory sync retries twice does not have to move with them. Under MIT, self-hosted, it doesn't leave your infrastructure at all — the same argument we made about context graphs being the real moat.
Shared memory means shared wrong answers. A single agent's bad inference dies with the session. A promoted one propagates to everybody's agent silently. Before you turn this on, decide who can promote a memory to team, and how a bad one gets retracted. Neither is in the box.
This is a lower-effort win than better prompts. We've already seen that coordination beats model size on real repos. CodeGraph is that idea persisted to disk: index the call graph once, and every agent starts from the map instead of grepping its way to one.
Read the license and the host, separately. MIT and self-hosted are exactly what you want. But this is a cloud vendor's open-source project — check what the default .env points at before you run it on a repo you care about.
Key takeaways
- TencentDB Agent Memory v2.0.0 shipped Aug 3, 2026 — MIT-licensed, self-hosted, one-command Docker deploy
- Four memory assets: Chat Memory, Skill, Wiki, and CodeGraph (symbols, files, call relationships)
- Three visibility levels — private, team, restricted (ACL) — with new assets private by default
- Team admins cannot read assets the owner hasn't shared
- No built-in workflow for correcting or retracting a shared memory that turns out to be wrong
- Decide your promotion and retraction rules before enabling team-wide sharing
Your team's hardest-won context shouldn't live in a chat history nobody can search. We stand up self-hosted agent memory and code graphs, with the promotion and retraction rules that keep a bad answer from spreading. See how we build agent infrastructure or tell us what your team keeps re-explaining.
Sources: TencentDB Agent Memory on GitHub, VentureBeat.
- #agent-memory
- #ai-agents
- #coding-agents
- #open-source
- #developer-tools
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Torvalds: huge AI-reviewed kernel RCs are the new normal
Linus Torvalds says AI review tools have permanently inflated Linux release candidates. The lesson for your team is about review capacity, not AI.
Read itDynamoDB vector search GA: one less database to run
DynamoDB vector search is generally available — embeddings stored next to operational data, single-digit ms at 99%+ recall. What it removes and what it can't filter.
Read it