TencentDB Agent Memory Turns Chat Histories Into Governable Team Assets

Agent memory products usually stop at vector search over chat logs. This system instead treats agent experience as versioned, access-controlled, role-assembled assets—changing the boundary from "what can the model recall" to "what should this specific agent be allowed to know." The 14k-star traction in four months signals real demand for memory architectures that go beyond chunk-and-embed, but the missing tests and Beta label mean production adoption still requires its own validation.

Instead of just chunking conversations for vector search, TencentDB Agent Memory decomposes agent experience into four governed asset classes. Chat Memory captures facts, preferences, and decisions refined through an L0-to-L3 pipeline. Skills extract repeatable workflows with versioning and access control. LLM-Wiki restructures documents into linked, deterministic Markdown pages. CodeGraph indexes symbols and call relationships for structural queries, not another RAG layer.

A Memory Proxy sits between the agent client and upstream LLM, injecting the right assets at the right injection points based on team, agent, and task identity. It speaks both Anthropic and OpenAI protocols, with non-fatal hook failures so one broken knowledge source doesn't crash the main request. The pipeline runs on a competing-consumer worker model with distributed locks, dead-letter queues, and configurable backoff.

The repo has pulled 14,327 stars in four months and ships a substantial 176,000-line TypeScript codebase across four services. But the public default branch carries no test files, component versions are misaligned, and the Team Memory feature is still labeled Beta—making it a technically deep early platform rather than a turnkey knowledge appliance.