+1,223 stars this week (5.1k total), and **10 releases between Aug 21–29** (v1.30.0 → v1.35.0). Single Rust binary: git-versioned markdown wiki + SQLite FTS5 + optional embeddings, exposed over HTTP/MCP. Hooks into Claude Code, Codex, Cursor, Gemini CLI, OpenCode. v1.34.0 (Aug 28) added an embedding-only credential and a generic conversation-JSON importer.
+2,207 stars this week (2,651 total), top Python repo on weekly trending. Read-only mirror of the marketplace for Claude Cowork *and* Claude Code, synced nightly from Anthropic's internal review pipeline; submissions go through automated security scanning.
~27.7k stars. Claims **95.2% R@5 on LongMemEval-S** (500 questions) and ~92% fewer tokens vs full-context (~170K tokens/year vs 650K+). Ships "0 external DBs" — SQLite only — plus an MCP server with 54 tools, and a thin shim falling back to 7 core tools when no server is reachable.
+2,353 stars this week, 38.8k total — sustained, not a spike (launched May 12, 2026, hit trending day one). Local-first memory graph compressed to scored Markdown trees in SQLite, mirrored as an editable Obsidian vault. Model routing defaults to **Anthropic Claude**, with local Ollama and BYOK; connects 5,000+ MCP servers.
+1,918 stars this week (~4.0k total), #1 trending TypeScript. Accepted into the **Apache Incubator** on a vote called Aug 9, 2026; **v0.1.11 shipped Aug 18** with 375 merged PRs from 24 contributors. Model messages, tool calls, tool results, permission decisions and termination events all recorded append-only in local SQLite; BYO model via MCP.
[Aug 25 security release](https://nextjs.org/blog/august-2026-security-release). CVE-2026-75604 (Windows path traversal, Pages Router or App Router without Cache Components) and a libheif/sharp heap overflow in the Image Optimization API — an attacker-controlled AVIF writes ~16KB past the allocation ([GHSA-2xp9-vwfh-vxw4](https://github.com/vercel/next.js/security/advisories/GHSA-2xp9-vwfh-vxw4)). Affects 13.4–15.5.23 and 16.0–16.3.2. Patches ship with AVIF optimization **disabled**, pending upstream.
[TS 0.122.0 / Python 1.2.0, Aug 27](https://platform.claude.com/docs/en/release-notes/api) — `client.beta.files` and `client.beta.skills` stop sending beta headers and return GA shapes; `BetaSkill` renamed `BetaContainerSkill`; `beta.skills.delete()` now deletes all versions. Preceded by [0.121.0, Aug 26](https://github.com/anthropics/anthropic-sdk-typescript/releases): **Standard Schema support for structured outputs and tools**, plus tool runner continuing on `pause_turn`.
[Aug 26 advisories](https://github.com/honojs/hono/security/advisories): GHSA-crvj-82cr-hjcx (query params after the URL fragment cause cache-key/proxy divergence), GHSA-gqvv-2mrq-wpjv (incomplete fix for CVE-2026-39408 — `toSSG()` still writes outside the output dir), GHSA-g6gw-c38x-mqfc (unbounded dot-notation nesting in `parseBody()` → memory exhaustion).
[v0.33.1, Aug 26](https://github.com/ollama/ollama/releases) adds structured output to the MLX runner and MLX support for Qwen 3.8 Flash Next; v0.33.0 (Aug 21) enables Claude Desktop integration as a third-party gateway provider and adds KV-cache prefill restore points; v0.33.2 (Aug 27) fixes the Claude Desktop proxy interrupting in-flight requests on catalog update.
Aug 25 — `ctx.abort({ retryAlarm: false })` stops an aborted alarm from retrying after the DO resets. Aug 28 — a DO can now have **ten** Dynamic Workers with in-flight requests, up from four. Also `[email protected]` and Local Explorer 0.15.0 (email handler inspection) on Aug 28.
[HN #49445727](https://news.ycombinator.com/item?id=49445727) — 501 pts / 213 comments, Aug 26. Rafael Pierre argues full pre-embedding is right for only ~10% of systems (10K+ queries/day, stable corpus); ~60% should stop at BM25 + LLM query rewriting (~$0.001/query, no re-indexing), with hybrid + on-the-fly embedding only when proven. Core claim: chunking strategy and embedding-model deprecation are hidden costs vector DBs impose. Note the essay itself is dated June 10, 2026 — the *discussion* is this week's.
[HN #49466917](https://news.ycombinator.com/item?id=49466917) — 774 pts / 339 comments, Aug 27. Calvin French-Owen (Segment co-founder, ex-OpenAI) splits workloads into "IQ 180" novel reasoning vs "token spewer" execution, and cites a task dropping from ~$1 on Sonnet-class to ~$0.10 on gpt-5.6-luna. Puts GLM-5.3 on the cost/capability Pareto frontier.
[GLM-5.3-Flash, HN #49449507](https://news.ycombinator.com/item?id=49449507) — 1124 pts / 571 comments, Aug 26; [GLM-5.3 open-weight, HN #49479878](https://news.ycombinator.com/item?id=49479878) — 703 pts / 234, Aug 28. Flash is 320B total / 18B active, MIT-licensed, hybrid sparse+linear attention (34 of 45 layers linear) claiming 3.01x less attention compute and 4.44x smaller KV cache; reported near Opus-class on coding/agentic at ~1/10 price. [unsloth GGUFs](https://huggingface.co/unsloth/GLM-5.3-GGUF) landed Aug 28.
[HN #49461817](https://news.ycombinator.com/item?id=49461817) — 663 pts / 319 comments, Aug 27; also picked up by [Daring Fireball](https://daringfireball.net/linked/2026/08/27/the-load-bearing-vocabulary-of-claude). Louis Abraham clustered 461,121 GitHub PR descriptions (Jan 2025–Aug 2026) using k-means with KL divergence over word distributions. One style went from 0.7% → 39% of the corpus — the one that says "load-bearing." [Repo is public.](https://github.com/louisabraham/load-bearing)