TencentDB Agent Memory: A Local Memory Hub That Stops AI Agents From Forgetting

Tencent’s MIT-licensed memory hub runs on your own machines and turns conversations, docs, and code into reusable memory assets. On Tencent’s own tests, it cut an agent’s token use by 61% and lifted its task pass rate by 52%.
artificial-intelligence
software-engineering
Author

Kabui, Charles

Published

2026-08-28

Keywords

agent-memory, local-memory-hub, tencentdb-agent-memory, memory-assets, token-savings

Download as PDF

TencentDB Agent Memory is a self-hosted memory hub for AI agents, open-sourced by Tencent’s cloud database team under the MIT license. It turns the raw material an agent already works with into reusable memory assets. Chat Memory distills conversations through four layers, from raw record to facts and preferences, project scenarios, and a stable persona; Skill captures proven workflows; Wiki structures documents; and CodeGraph maps how code calls connect. Version 2.0, released August 3, adds a Memory Hub where teams create agents and review, share, and assign these assets under private, team, or role-based visibility. In Tencent’s own tests, an OpenClaw agent using the memory plugin cut token use by 61% on its WideSearch benchmark while lifting the pass rate from 33% to 50%.

Agents are billed per token and forget between sessions, so every time one re-explains your project or re-reads a file, you pay twice. Because it runs on your own machines, your data never leaves them, and a built-in proxy lets Claude Code, Codex, DeepSeek Harness, and other tools plug in by pointing at an address, no plugin needed. A solo developer can run a small team of agents that inherit each other’s experience instead of re-teaching them from scratch.

Agent memory is moving from an afterthought to a governed, shared asset, and the open-source versions run where you can see them. It is not a settled standard: Tencent says automatic memory routing is still being refined, and the benchmark numbers are its own.

Read More: see why bigger context windows alone don’t stop agents from forgetting.

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {TencentDB {Agent} {Memory:} {A} {Local} {Memory} {Hub} {That}
    {Stops} {AI} {Agents} {From} {Forgetting}},
  date = {2026-08-28},
  url = {https://toknow.ai/posts/tencentdb-agent-memory-local-memory-hub-agents/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “TencentDB Agent Memory: A Local Memory Hub That Stops AI Agents From Forgetting.” https://toknow.ai/posts/tencentdb-agent-memory-local-memory-hub-agents/.