Tencent open-sources a memory hub that lets AI agents share what they learn—without leaking secrets

AI agents forget everything the moment a session ends. That’s been the deal forever. You onboard a coding agent to your repo, explain the architecture, set some preferences, and tomorrow it’s amnesia all over again. Tencent Cloud just open-sourced something that actually fixes this.

The TencentDB Agent Memory Hub v2.0 turns conversations, documents, and code into four persistent memory assets: Chat Memory, Skill, LLM-Wiki, and CodeGraph. The clever part isn’t storage — it’s governance. An agent on your teammate’s machine can read what your agent learned, but only if you’ve explicitly shared it. Private memories stay private, even from team admins. New chat memories and skills default to private, making sharing an intentional act rather than a data leak waiting to happen.

Under the hood, chat conversations go through a layered distillation pipeline. Raw logs (L0) get refined asynchronously into atomic facts (L1), scenarios (L2), and core persona traits (L3). When an agent starts a new session, it bootstraps from L2 and L3 for fast context, then falls back to BM25 plus vector retrieval across L1 and L0 when it needs specific facts — all capped by item count, character budget, and timeout so memory doesn’t eat the entire context window. The Memory Proxy speaks both Anthropic and OpenAI protocols, meaning Claude Code and CodeBuddy can plug in without adapter code.

Tencent reports PersonaMem accuracy jumping from 48% to 76%, a 59% relative gain. That’s self-reported with no independent reproduction yet, so take it with the appropriate grain of salt. Still, the architecture is sound and the MIT license means you’re not locking yourself into a vendor. Solo developers and small teams get the most value right now; large regulated enterprises should pilot rather than standardize, since private-repo CodeGraph and automated memory routing are still being refined. If you’ve ever wished your coding agent remembered that you hate snake_case, this is worth a Saturday morning Docker pull.