[Preview launched Aug 3, 2026](https://blog.cloudflare.com/cloudflare-computer/); +6,020 stars in 7d (7.9k total), #1 on weekly trending for TypeScript and overall. A Durable Object-backed SQLite virtual FS shared across three backends: FUSE-mounted Linux container, isolate bash shell, and isolate JS in a Dynamic Worker. MIT, explicitly preview-not-production.
[v2.0.0 released Aug 3, 2026](https://github.com/TencentCloud/TencentDB-Agent-Memory); +5,720 stars in 7d (21k total). Four memory assets (Chat Memory L0–L3, Skill, LLM-Wiki, Code-Graph), a team governance layer for private-vs-shared memory, and [claimed 33–61% token reduction](https://www.marktechpost.com/2026/08/07/tencent-cloud-open-sources-tencentdb-agent-memory-v2-0/) by querying the persona tier before raw conversation. Self-hosted via Docker, MIT.
+3,585 stars in 7d (6.0k total), [hit #1 on GitHub Trending Aug 10, 2026](https://trendshift.io/repositories/18986). Pluggable backends across Neo4j / FalkorDB / Apache AGE and FAISS / Qdrant / **pgvector**; ships an MCP server for Claude and Cursor; exports provenance as W3C PROV-O. Positioned as an open-source Palantir alternative.
+1,511 stars in 7d (4.1k total). Memgraph for the call/symbol graph, Qdrant for semantic search, data-flow tracing across 10 languages, ast-grep structural search/replace. Runs as an MCP server so Claude Code queries and edits the graph directly.
[API release notes, Aug 7 2026](https://platform.claude.com/docs/en/release-notes/api) — sessions now take a hard spend cap that pauses with a `budget_reached` stop reason (resumable by raising the budget); a `{"type": "advisor"}` roster entry lets the primary thread consult a peer-or-better model mid-turn; and a mounted GitHub repo's root `.claude/skills` is discovered automatically at session start. Shipped same day in `anthropic-sdk-typescript` [v0.116.0](https://github.com/anthropics/anthropic-sdk-typescript/releases) and `anthropic-sdk-python` v0.121.0, alongside the `mid-conversation-tool-changes-2026-07-01` beta and `inference_geo` pinning.
[[email protected]](https://github.com/cloudflare/workers-sdk/releases/tag/wrangler%404.122.0) (Aug 12): "As of compatibility date `2026-08-04`, workerd enables the `nodejs_compat` and `nodejs_compat_v2` compatibility flags by default." Wrangler, `@cloudflare/[email protected]`, and `[email protected]` now infer Node compat from the date and strip the redundant flag. Opting out on a new date requires *both* `no_nodejs_compat` and `no_nodejs_compat_v2`. (The runtime change itself landed [Aug 4](https://developers.cloudflare.com/changelog/post/2026-08-04-nodejs-compat-default/) — 9 days ago; the SDK alignment is this week.)
Meta released [Muse Glimmer](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model), a 30B open-weights agentic model under Apache 2.0, on Aug 10. Ollama [v0.32.7](https://github.com/ollama/ollama/releases) (Aug 10) added MLX-engine support on Apple Silicon with DFlash speculative decoding and image input; v0.32.8 (Aug 10) extended it to NVIDIA/AMD; v0.32.9 (Aug 11) also added NVIDIA Nemotron 3.5 Lightning (30B MoE, 3B active). Release notes explicitly target "long-running personal assistants." HF ships BF16, GGUF k-quants, ExecuTorch, and the DFlash drafter.
[API release notes, Aug 10 2026](https://platform.claude.com/docs/en/release-notes/api): Sonnet 5's $2/$10 per MTok introductory rate is now standard — the increase to $3/$15 scheduled for Sept 1, 2026 will not happen.
[blog.cloudflare.com/mcp-v2](https://blog.cloudflare.com/mcp-v2/) (Aug 6). The new spec removes the init handshake, `Mcp-Session-Id`, and protocol sessions from the request path; deprecates HTTP+SSE (12-month window), Dynamic Client Registration, and Roots/Sampling/Logging. Elicitation is redesigned as Multi Round-Trip Requests (`input_required` + client retry). New required `Mcp-Method`/`Mcp-Name` headers. Sentry and Linear already in prod.
[Agents Week review](https://blog.cloudflare.com/agents-week-in-review/) lists Agent Memory alongside AI Search, Sandboxes GA, Artifacts, and Durable Object Facets. The [product post](https://blog.cloudflare.com/introducing-agent-memory/) (originally Apr 17, private beta) details the design: ingest/remember/recall/forget/list; DO+SQLite for storage, Vectorize for semantic search; classification into facts/events/instructions/tasks with supersession chains; five-channel retrieval (FTS, fact-key, raw message, direct vectors, HyDE) merged via Reciprocal Rank Fusion.
HN #1 this week, 1199 pts / 637 comments (Aug 10). 30B, Apache 2.0, ~20GB at 4-bit — runs on 24–32GB Macs. Explicitly optimized for tool invocation with precise schemas, multi-step recovery, and long-running local workflows; shipping via Ollama and LM Studio. Beats Gemma4-31B and Qwen3.6-27B on agentic benchmarks.
[arXiv 2608.09867](https://arxiv.org/abs/2608.09867) (Aug 10), HN 685 pts / 301 comments. Encrypted CoT blocks are interchangeable across sessions, users, and models within one provider; injecting a strong model's trace into a weaker sibling makes it emit plaintext. Demonstrated on OpenAI, Anthropic, and Google. From publicly scraped blocks the authors recovered 367 PII artifacts and 182 credentials. Disclosed May 2026; providers reportedly disputed the side-channel/replay framing.
HN 933 pts / 378 comments (Aug 12). GA MoE release, 1M context, roughly an order of magnitude under frontier pricing. Landed the same week as Qwen3.8-2.4T (631 pts) and DeepSeek V4 Flash matching March frontier scores on gaming PCs.