Hy-Memory: Tencent's open 30B-A3B memory writer, served locally with vLLM and tested on a Seattle-to-Sydney fact update

What the video shows

• Two pieces. The model is Hy-Memory-30B-A3B (about 3 billion active parameters), the "writer" that decides what to keep and how to update it. The SDK is the "notebook" it writes into. He installs both.
• The machine. Ubuntu with one 80 GB H100. The model is already served with vLLM and holds about 75 GB, mostly because of the maximum model length; lowering that setting lowers memory use.
• Install and connect. The SDK installs with one uv pip command. He then points the SDK's writer at the local vLLM server, and uses an Alibaba DashScope embedding API for vectors because it is quicker to set up, noting a local embedding model also works.
• The test. A short script creates a client in pro mode, the mode that uses a model to extract and reconcile facts. It adds "Alex lives in Seattle" and a job fact, then adds "Alex moved to Sydney", then searches where Alex lives.
• The result. The model does not stack a second note on top. It marks the Seattle memory as superseded and writes its reason ("the user moved from Seattle to Sydney and no longer lives in Seattle"), updates the job memory to carry the history, and the resulting profile says Sydney. SQLite is the store behind it.
• Benchmarks and training. He shows Tencent's chart: 55.49 overall for Hy-Memory against Claude Opus 5, DeepSeek V4.1 and a size-matched Qwen at 50.37, with only the memory-writing model changed. Training has two stages: a student writes memories while a frozen teacher checks each step, then reinforcement learning scores the whole memory bank after each chunk.

Checked against the official README

From the Tencent-Hunyuan/Hy-Memory README, read 2026-10-22.

• Install. pip install hy-memory==2.0.4 is the pinned command, with extras such as "hy-memory[qdrant]==2.0.4", [faiss], [graph] (Neo4j), [mysql], [rerank] and [all]; the core install includes ChromaDB and Kuzu. A newer 2.0.5 release appeared on the repository the same day, so check which version you want.
• Configuration. Only an LLM endpoint and an embedding endpoint are needed, and any OpenAI-compatible service works. They are set with MEMORY_LLM_API_KEY, MEMORY_LLM_BASE_URL and MEMORY_LLM_MODEL, plus MEMORY_EMBEDDER_API_KEY, MEMORY_EMBEDDER_BASE_URL, MEMORY_EMBEDDER_MODEL and MEMORY_EMBEDDING_DIMS. Pointing MEMORY_LLM_BASE_URL at a local vLLM server is how the video does it, although the video does not show the variable names.
• The API. HyMemoryClient(mode="pro"), then client.add(...) with text or OpenAI-format messages and a user_id, and client.search(...), which returns profile, normal and proactive channels. The README says writes for the working and coding scenes need a tool-call trace and only take effect in pro or ultra mode; lite stores everything as raw memory.
• Smaller builds. The README says FP8 and GGUF versions of the model were released in October. A third-party report put the original BF16 download at about 60 GB, so the 80 GB card in the video is not the minimum.
• Licence. Apache-2.0 for the repository and the model page. The README also describes MemHub, a local-first desktop and command-line app that gathers history from tools such as Claude Code, Cursor and Codex.

Gotchas & Caveats

• The benchmark figures are Tencent's own and were not independently reproduced.
• His test is one fact change. It shows the supersede behaviour working, not how the model handles long, messy histories.
• He says he plans to use it in production with a client; that is his judgement, not a result we can check.

Key Takeaways

• Test a memory system by changing a fact and searching for the new one; the old one should be marked superseded, not left beside it.
• Serve the writer on an OpenAI-compatible endpoint and set the SDK's environment variables to point at it.
• Try the FP8 or GGUF build before renting a large GPU.