On Oct 16 Tencent Hunyuan open-sourced Hy Memory 2.0 in three layers: the Apache-2.0 memory-writer model Hy-Memory-30B-A3B (~3B active, HF weights plus arXiv report), a Python SDK covering Chat, Working and Coding scenes (PyPI hy-memory 2.0.x), and local-first MemHub (CLI + macOS, preview). Per Tencent, swapping it in as the writer across Mem0, MemOS, Hy-Memory and held-out EverMemOS yields a 55.49 average vs 54.62 for Claude Opus 5, measured on fixed 20% subsamples; it still trails Opus 5 on several individual metrics.
Key Takeaways
- ✓Vendor-reported: 55.49 four-harness average vs 54.62 for Claude Opus 5 and 53.64 for DeepSeek V4.1 Flash
- ✓Held-out EverMemOS: 56.89 vs Opus 5 at 54.04; still trails on Mem0 LoCoMo (59.54 vs 69.47)
- ✓~3B active, Apache-2.0, tested on vLLM v0.31.0 with an OpenAI-compatible API
- ✓SDK installs via pip install hy-memory with automatic Chat/Working/Coding routing
- ✓MemHub runs on Linux and macOS (preview); Windows unsupported
Key Decision Metrics at a Glance
Heavy Claude Code use: compare subscription limits and API bills
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Core Background & Industry Pain Points
Agent memory frameworks such as Mem0 and MemOS usually delegate "what to store and how to update it" to a general LLM: flagship models are costly, small ones miss or corrupt facts. Coding is worse, since Claude Code, Cursor and Codex each keep separate context.
Architecture Highlights & Internals
Hy Memory 2.0 ships three layers. Hy-Memory-30B-A3B (~30B total, ~3B active) handles only the memory write path and drops into those frameworks in place of the general LLM; the report describes teacher-supervised distillation (OPD) inside deployed harnesses followed by chunk-level RL. The SDK offers Chat, Working and Coding scenes (Coding adds a version-evolution diff chain) with automatic routing. MemHub aggregates user-authorized chats and coding traces locally and grants them to each tool separately.
Authoritative Benchmarks & Measured Scores
All figures are Tencent-reported on fixed 20% subsamples. Four-harness average: 55.49 vs Claude Opus 5 54.62, DeepSeek V4.1 Flash 53.64, Qwen3.6-35B-A3B 50.37. On held-out EverMemOS: 56.89 vs 54.04. The lead is under one point, and it still trails Opus 5 on Mem0 LoCoMo (59.54 vs 69.47). No third-party reproduction yet.
Developer Hands-on Guide
Teams on Mem0/MemOS can serve it with vLLM (tested v0.31.0) as hy-memory, swap the writer, and compare cost and recall. Install the SDK with pip install hy-memory. MemHub is a preview for Linux x64/arm64 and macOS, not Windows; check release assets and the privacy notes first.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.