[github.com/denoland/celld](https://github.com/denoland/celld) — +935 stars this week, still on Rust weekly trending at 3.8k total. v0.1.0 shipped Aug 5; Apache-2.0. Runs Workers + DO on your own boxes, each cell = V8 + SQLite, coordination via an S3/GCS bucket only — no control plane, no consensus.
[github.com/TencentCloud/TencentDB-Agent-Memory](https://github.com/TencentCloud/TencentDB-Agent-Memory) — +3,637 stars this week (22.5k total), #1 TS trending after prime-agent. MIT. v2.0.0 out. Layered L0 conversation → L1 atom → L2 scenario → L3 persona, with vector + BM25 fused by RRF, plus CodeGraph and Wiki assets.
[github.com/vitali87/code-graph-rag](https://github.com/vitali87/code-graph-rag) — +1,686 stars this week (4.5k total). Tree-sitter parsing across 12+ languages into a Memgraph property graph, exposed to Claude Code via MCP; recent additions are runtime call tracing and structural search/replace.
[github.com/NVIDIA-NeMo/Switchyard](https://github.com/NVIDIA-NeMo/Switchyard) — +1,435 stars this week off a 1.75k base (near-total growth this week). Rust proxy accepting OpenAI Chat/Responses *and* Anthropic Messages, with escalation routing (weak tier first, judge, escalate) and Ollama/vLLM backends. Self-labeled pre-alpha.
[github.com/semantica-agi/semantica](https://github.com/semantica-agi/semantica) — +5,284 stars this week (8.4k total), #1 Python trending. MIT, Python. Backends include Neo4j *and* pgvector; W3C PROV-O provenance per fact, conflict detection, point-in-time snapshots, and an MCP server. Caveat on signal: earlier Show HN attempts under a prior org (Hawksight-AI) scored 2–8 points, so this week's spike lacks HN corroboration.
Python `0.122.0` and TypeScript `0.117.0`, both Aug 13, 2026: "add `output_behavior` to dream creation — create new memory store or update the input store in place." This inverts the documented Dreams invariant ("The input store is never modified" / copy-on-write). Same release also fixes `message_delta` accumulation on streamed messages and makes the tool runner forward container IDs.
Aug 14 changelog: Access policies now bind to the Worker itself, so every route, Custom Domain, and preview URL stays protected as routes change; plus an account-wide "all Workers private by default" mode with per-Worker bypass. Wrangler 4.123.0 (Aug 13) adds mock identities in `wrangler.json` so `ctx.access.getIdentity()` works in local dev.
C3 2.72.0 (Aug 13) switches the default Next.js-on-Workers scaffold to vinext, Cloudflare's Vite plugin reimplementing the Next.js API surface (~94% of Next.js 16, 4.4× faster builds, 57% smaller bundles per Cloudflare). This displaces OpenNext as the paved path.
v0.32.12 (Aug 14) adds `ollama run qwen3.8:27b` — a 27B dense native vision-language model, Apache 2.0, with MLX perf work for Apple Silicon. Reported to more than triple its predecessor on DeepSWE (13.3 → 42.2). v0.32.14 (Aug 15) adds WebP transcoding; v0.32.13 adds developer-instruction support.
Aug 12 changelog entry: Gateway automatically identifies Model Context Protocol traffic, exposes new policy selectors for it, and adds an "AI security report" dashboard showing MCP usage. (Detail page returned 404 — headline is from the changelog index.)
[research.meta.ai](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) (Aug 10) — 1,205 pts / 638 comments on HN. 30B, Apache 2.0, K-Quant 17GB fits 24–32GB consumer GPUs; explicitly trained for multi-step tool-calling, tool-failure recovery, and vision. Benchmarked vs Gemma4-31B and Qwen3.6-27B.
[simonwillison.net](https://simonwillison.net/2026/Aug/16/qwen-38-27b/) (Aug 16, 504 pts); model thread Aug 14 at 1,419 pts / 788 comments. Apache 2, 262k context, 17GB Q4_K_M, 15–30 tok/s on M5 Max. Simon's core finding: it defaults to `xhigh` reasoning and burns minutes on trivial requests — use `low` or none for real work.
[docker.com](https://www.docker.com/products/docker-sandboxes/) (Aug 10) — 693 pts / 396 comments. Comment thread is the real signal: microVM-vs-container isolation, TLS-interception weaknesses, and the point that VM isolation alone doesn't stop credential/SSH-key exfiltration. Compared throughout to Podman, Bubblewrap, credential brokers.
[platform.claude.com](https://platform.claude.com/docs/en/release-notes/system-prompts) (HN Aug 16, 665 pts / 256 comments). Full text for Opus 5 (Jul 24), Fable 5, Opus 4.8, Opus 4.7, Sonnet 4.6 — plus the note that API users get a *different or no* system prompt.
[mixedbread.com](https://www.mixedbread.com/blog/toast-1) (Aug 13, 219 pts / 66 comments). Not an embedding model — a search agent that decomposes queries and curates context. Claims 7–11× cheaper than frontier models on BrowseComp Plus / OfficeQA Pro; 3.5× token reduction on Harvey LAB at parity. Backend-agnostic, no migration. $0.30/M in.