The 'Vector Databases Are Dead' Claim Came Too Early—but Tencent's Open-Source Memory Layer Has Real Substance

A tweet says China has killed the entire vector database industry.

The reason is TencentDB Agent Memory, open-sourced not long ago, which adds long-term memory to any AI agent and runs entirely locally. No Pinecone, no cloud API, no need to introduce yourself again every time you start a new session.

First, let's look at what it actually does.

Most memory systems compress your history into an opaque blob of vectors; when recall goes wrong, you can only guess which step caused the problem. This project does the opposite: it doesn't compress, it layers. L0 raw conversation, L1 atomic facts, L2 scenarios, L3 persona profiles. Short-term state is compiled into a Mermaid diagram in the agent context, and verbose tool logs are dumped to disk. When evidence is needed, you drill back along the node_id to reach that original log. Every layer is markdown you can open and read directly.

The numbers are direct too: 61% fewer tokens, PersonaMem accuracy from 48% to 76%, running on plain SQLite with no external API dependencies. The original post says 100% open source, 5.1k stars.

Then comes the cold water.

The first reply immediately takes apart '100% local': what is local is the SQLite storage part. One reply mentions that its Hermes image by default sends memory extraction to DeepSeek-V3.2 on Tencent Cloud, so chat content leaves your machine unless you point MODEL_BASE_URL to Ollama. Also, if you start the service with -p 8420:8420 and don't set TDAI_GATEWAY_API_KEY, the recall endpoint is left exposed to your LAN.

The second bucket of cold water is even more important: the README says BM25 plus vector retrieval. The vector database is not dead; it has just been hidden in the pipeline where users cannot see it.

Some people also nitpick the 76% figure and ask for the source of the benchmark. That's not an unreasonable request.

My take: using 'kills vector databases' as a headline is clickbait; taking it seriously would be a mistake. What is really rejected is the lazy approach of shoving whatever content into a vector pile, with inaccurate recall and no way to troubleshoot. Layered, readable, and traceable back to the original logs—this approach is a real inspiration for people building agents.

But 'local' and 'private' are two different things. Before deploying, look at the default configuration first, keep model requests on your local machine, and lock down ports and keys. If you want to try it, first set the two environment variables MODEL_BASE_URL and TDAI_GATEWAY_API_KEY correctly, and then talk about memory.