--- title: "Approaching Human Memory — Tencent's Agent Memory" date: 2026-10-08 time: 19:10 model: admin category: reviews summary: A team-level agent memory hub Tencent Cloud released under the MIT license. It turns conversations, docs, and code into four memory assets and manages access permissions too. Install path and how it's actually used. tags: TencentDB,Agent Memory,agent,memory,team collaboration,MIT,open-source --- # Approaching Human Memory — Tencent's Agent Memory Explaining project background to an agent over and over is exhausting. Docs get read from page one every time, and whatever approach worked in the last session vanishes in the next. TencentDB Agent Memory is the open-source project that set out to kill that repetition. Tencent Cloud's database team published it under the MIT license, and the repo name is TencentCloud/TencentDB-Agent-Memory. Rather than a tool that simply leaves conversations in memory, it's closer to a hub that manages memory assets at the team level and hands them out to agents. --- ## What it's trying to solve The starting question this project took is a practical one. If you've already explained the project context, don't you not need to explain it again in a new session? If the docs have been read, does every agent need to reread from the first page? Do you really need to flounder again over a workflow that already worked once? The flow is: turn existing information into reusable memory assets, and that leads to fewer turns and less rework. The metaphor the README uses is a save file. Most agents' first mission is relearning the project, and this tool turns that already-paid learning cost into a save file the whole team can load from the start. --- ## Four memory assets It doesn't keep memory as one lumped blob. It splits it into four assets. - **Chat Memory** — preferences, facts, decisions, and interaction history. No need to reintroduce yourself every time you spin up a new agent. - **Skill** — repeatable work procedures. Includes version, resource files, trigger boundaries, execution steps, and verification rules. Not a prompt snippet you paste in, closer to a version-controlled execution manual. A code review checklist or an incident-response runbook — learn it once and the whole team uses it. - **LLM-Wiki** — product docs, design specs, and ops runbooks turned into structured pages with a link graph. It states it was influenced by Karpathy's LLM knowledge base idea. - **Code-Graph** — indexes code symbols, files, call relationships, and impact paths. Where simple RAG tells an agent where the code is, CodeGraph tells it that changing this spot can affect that one over there. You can check callers and callees before touching a function and do impact analysis. The Skill mechanism is borrowed from Nous Research's Hermes Agent, and CodeGraph is based on the open-source codegraph project. --- ## A human-like four-stage memory hierarchy Chat Memory isn't left as a flat log. Once the raw conversation is stored, an async pipeline condenses it in stages. | Stage | What's stored | What it's for | | --- | --- | --- | | L0 Conversation | Raw conversation logs with full context | Checking expressions, timestamps, sources | | L1 Atom | Extracted facts, preferences, constraints, events | Precise retrieval of actionable info | | L2 Scenario | Project- and scenario-centric knowledge blocks | Rapid context restoration in a new session | | L3 Core / Persona | Long-term profile, stable behavior patterns | Instant sync with the team's core context | Retrieval is hierarchical too. Usually L2 and L3 grab the context quickly first, and you only descend to L1 and L0 when a specific fact is needed. Search is hybrid — BM25, vector search, and RRF (Reciprocal Rank Fusion) combined. Results are capped by item count, character budget, and timeout, so memory can't crowd the actual work out of the context window. A direct response to the failure mode where you bolt on memory and the prompt fills with nothing but memory. --- ## Integration and access control Integration is a single proxy server. No need to build separate plugins, hooks, or MCP servers — just point the agent's Base URL at the proxy. The official support list includes Claude Code, Codex, CodeBuddy, WorkBuddy, Hermes, OpenClaw, and DeepSeek Harness. SDKs are provided for TypeScript and Python. Installation spins up three Docker images. memory-core, memory-hub, and proxy come up with one script, and the panel opens at localhost:8125. Default ports: core 8420, panel 8125, knowledge 8424, proxy 8096. It's self-hosted, and the MongoDB storage backend is experimental so it's off by default. Access control is what most evaluations call this project's real differentiator. RAG asks what can be found; team memory asks who can use it, which version is valid, and which agent gets it. Visibility comes in four levels. - **private** — owner only. Not even the team admin can read it. - **team** — whole team reads; only owner and admin manage. - **restricted** — fine-grained control by user, role, and agent ACL. - **agent** — handed only to specific agents within the team. New Chat Memory and Skill default to private. Sharing is an explicit act. For a tool handling real codebases, that default is the right one. You can equip things like: release Skill to the release agent only, architecture Wiki to all the dev agents, and CodeGraph to coders and reviewers only. --- ## Numbers and performance claims A summary of the figures from the official announcement and repo docs. All the numbers below are the maker's claims. - Stable v2.0.0 released August 3, 2026, followed by v2.0.1 (August 25). - TeamMemory update announced August 13, 2026, extending personal memory to team collaboration. - Per the official announcement, it passed 20,000 GitHub stars in 90 days and hit #1 on GitHub Trending. As of October 8, 2026 the repo shows about 27,788 stars. - Internal benchmarks claim up to 61.38% token savings and a relative 51.52% improvement in task pass rate. - Self-reported PersonaMem rise from 48% to 76%. No independent reproduction yet. You don't have to take the numbers at face value. The direction is clear enough though: cut the memory and set a budget, and the context shrinks; shrink the context and cost and errors drop with it. That's the design proposition. --- ## Compared to competing tools Single-agent memory tools are already plentiful — Cognee, Mem0, Supermemory, Honcho, and others. Where this project takes a different approach is governance. Even if a colleague's agent can read what my agent learned, it can only do so when I've shared it. Default is private, and the ACL genuinely works. If it's one person, one agent, and persistent memory, other tools can be simpler. If several people are running several agents and there's information not everyone should see, an open-source option that treats this as a first-class concern is still rare. --- ## Verdict TencentDB Agent Memory has a clear direction in that it treats memory as a team asset rather than a session byproduct. The four-asset structure, hierarchical condensation, budgeted hybrid search, and a management panel that ties owner, version, and ACL together all flow as one. It's MIT-licensed and self-hosted, so the dependency burden isn't heavy. There are still things to check. The benchmarks lean on self-reported figures. TeamMemory moves fast, so docs and behavior keep changing. And because it's a proxy-based structure, you have to gather your existing agent configs in one place. If you've started running AI agents as a team, this is worth a read to check whether the time has come to manage memory as a team asset rather than an individual one. Deciding upfront how to divide up memory assets and who gets what tends to be something better settled before the code.