Tencent open-sources Hy Memory 2.0, a memory stack for your agents
An Apache-2.0 SDK, a 30B memory model and a local hub that pools recall across Claude Code, Cursor and Codex — plus a new open coding model and more.
Copy markdown
A 30B memory model that extracts and reconciles
Hy-Memory-30B-A3B (~3B active) scores 72.6% on LoCoMo while cutting tokens per question from 1,159 to 703 — Tencent claims it beats Claude Opus 5 at long-term recall. Apache-2.0 weights are on Hugging Face; the SDK installs with pip install hy-memory==2.0.1.
MemHub pools memory across Claude Code, Cursor and Codex
A local-first desktop and CLI app unifies conversation history and coding traces from Claude Code, Cursor, Codex, OpenCode and Qwen Code into one local runtime and data directory. Your agents stop forgetting between tools and sessions — and the memory never leaves your machine.
Build this weekend: cross-session agent recall
Point the SDK at any OpenAI-compatible LLM and embedding endpoint, set scene="coding" or "working", then call add() and search(). It ships Chroma, Qdrant and FAISS vector stores plus Kuzu and Neo4j graph backends, so a coding agent that remembers past fixes and the diffs behind them is a day's work.
Ant's Ling 3.1 Flash opens its weights under MIT
InclusionAI published the full Ling-3.1-flash release: a 560B MoE with 25B active, up to 1M context, MIT-licensed fp8 weights, and day-0 SGLang and vLLM recipes plus a LingCode harness (self-reported SWE-bench Pro ~68.7). It reportedly fits on four GPUs — another frontier-class coder you can self-host for free.
Anthropic reveals Scion, its in-house git replacement
After merges hit 30,000 a day on a ~100GB monorepo, Anthropic swapped git for Scion — server-side changesets and virtualized checkouts that cut infra CPU and memory ~200x and checkouts from minutes to seconds. It won't ship publicly, but the write-up is a preview of where version control for always-online agents is heading.