A recurring annoyance with AI coding agents is that every new session starts from zero: the agent has no memory of decisions made, conventions established, or context explained in a previous session, let alone one run by a different teammate. Tencent Cloud's newly open-sourced TencentDB Agent Memory addresses this by turning conversations, documents, and code into four distinct, reusable memory types: a running log of decisions and preferences, distilled skills extracted from completed tasks, a wiki-style knowledge base built from project docs, and a code graph that tracks symbols, files, and the call relationships between them. What sets it apart from a typical retrieval-augmented-generation setup is less the retrieval mechanism, which layers keyword search, vector search, and reciprocal rank fusion, with older conversation turns automatically distilled into progressively more compact summaries, and more the access-control layer sitting on top of it: memories can be marked private, shared with a team, or restricted, and one teammate's agent can read what another agent already learned about a codebase without exposing anything marked private. That is a real gap in most current agent-memory tooling, which tends to assume a single user working with a single agent rather than a team of engineers whose agents need to share institutional knowledge without a free-for-all. The project is MIT-licensed, ships as a handful of Docker containers that start with one command, and integrates with Claude Code, among other agent runtimes, via a proxy that speaks both the Anthropic and OpenAI message formats and automatically injects the right memory into a session's system prompt before forwarding the request. Tencent reports a jump in a memory-recall benchmark from 48 percent to 76 percent accuracy, though that figure is self-reported and has not been independently verified. For teams running multiple coding agents across a shared codebase, this is a concrete answer to a problem most agent frameworks have left to ad hoc solutions: how do you let agents build on each other's work without repeating the same onboarding explanation every single session.