TencentDB Agent Memory gives a whole team of AI agents a shared, persistent memory for free, the kind of infrastructure Mem0 and Zep charge $249 to $375 a month to run.
Tencent just put a piece of infrastructure on GitHub for free that two venture-backed startups currently charge production teams $249 to $375 a month to run. It's called TencentDB Agent Memory, and what it does is give an entire team of AI agents, not just one chatbot, a shared memory of your company's decisions, code, and conversations that persists across sessions and across which framework the agent happens to be running on.
Mem0's Pro tier, the one built for "production apps with real scale," runs $249 a month once you're past the free tier's 1,000 monthly retrieval requests. Zep's comparable Flex Plus plan runs $312 to $375 a month depending on billing cycle. TencentDB Agent Memory does a similar job, self-hosted, MIT licensed, at whatever your server costs.
What changed. The repo hit 13.2k stars and gained roughly 1,100 of them in the last day, timed to a v2.0.0 release on August 3. It ships as three services (memory-core, memory-hub, and a proxy) you start with one script, and it builds four kinds of memory a team's agents can share: Chat Memory (preferences, facts, and decisions from past conversations), Skill (reusable procedures extracted from completed work, versioned like code), Wiki (docs and specs turned into structured, linked pages), and CodeGraph (an index of a codebase's symbols and call relationships, so an agent can check what a change might break before touching it). All four get bound to specific agents through a permissions panel, so a "Scout" agent gets market research memory, a "Reviewer" agent gets an incident history, and neither one drowns in context it doesn't need.
The company backs this with a benchmark most teams will recognize the shape of: on PersonaMem, a test of whether an agent correctly recalls user context after a long interaction, accuracy went from 48% to 76% with the memory layer turned on. On OpenClaw integration tests, token usage dropped by up to 61% on long-horizon tasks while task success rates went up.
Who it's for. This is built for a team already running more than one agent, whether that's a two-person startup's "Scout, Builder, Reviewer" setup or an internal platform team standardizing memory across Claude Code, CodeBuddy, and custom SDK agents. The stated integrations today are OpenClaw, Hermes, Claude Code, and CodeBuddy, so if your stack lives outside that list, you're waiting on a roadmap item, not a supported path. A solo founder juggling one agent doesn't need this yet; a team onboarding a fourth or fifth agent that keeps re-learning the same project from scratch is exactly the buyer this replaces.
The honest caveat. Free doesn't mean zero cost. You're standing up and operating three services yourself, which is real infrastructure time even with a one-command installer, and that's before you account for the LLM API calls the memory pipeline itself makes to extract and summarize what it stores. The project's own notes admit automated memory routing is still "under iteration," meaning a human is still doing some of the binding work by hand, and CodeGraph currently prioritizes public HTTPS repos, with private-repo and SSH support still being refined. There are two listed contributors on the repo, which is thin for something you'd bet a production workflow on without a fallback plan. Mem0 and Zep charge for exactly the parts this project asks you to run yourself: the managed uptime, the SLA, the support Slack channel when recall breaks at 2 a.m.
The pitch here isn't that self-hosting beats a subscription on every axis. It's that the axis where TencentDB Agent Memory wins, teaching a whole team of agents to stop re-learning the same context every session, used to require either building it in-house or paying an enterprise rate for someone else's version of it. Now it's a git clone and an afternoon.