[Show HN, Aug 25 2026](https://news.ycombinator.com/item?id=49437049), 187 pts / 54 comments; 591 stars, MIT, written in Zig with zero deps. Ships Python **and** TypeScript/Node bindings (`npm install @hajewski/latticedb`). Claims 0.13 μs node lookups, 0.83 ms vector search @ 1M vectors at 100% recall, BM25 ~300x faster than FTS5. Author shipped two releases in 24h off HN feedback.
5.2k stars, +1,086 this week (Rust trending #8). Four releases Aug 28–29 (v1.33.0 → v1.36.0) adding a standalone importer and `workstreams` command. Storage is git-versioned Markdown as source of truth + SQLite for FTS5/embeddings. Wired into 20+ agent CLIs via MCP.
2.8k stars, +2,234 this week (#1 Python trending); companion `claude-plugins-official` +1,603. Read-only mirror syncing nightly from Anthropic's internal review pipeline; submissions go through [clau.de/plugin-directory-submission](https://clau.de/plugin-directory-submission) and must pass automated security scanning. Direct PRs are auto-closed.
34.4k stars, +2,326 this week; ByteDance-backed. Just announced Managed SaaS and Self-Managed commercial editions. AGPLv3 core (Apache-2.0 CLI). Models memory/resources/skills as a virtual filesystem under `viking://` with tiered L0/L1/L2 loading. *Freshness caveat: open-sourced March 2026 — the commercialization, not the launch, is this week's signal.*
[GHSA-2xp9-vwfh-vxw4](https://github.com/vercel/next.js/security/advisories/GHSA-2xp9-vwfh-vxw4) (Aug 25, CVSS 9.5) — unauthenticated RCE via the Image Optimization API on AVIF files, rooted in `libheif` under `sharp`; AVIF optimization is now disabled pending upstream fix. Same day: GHSA-p293-qw3h-jr36, unauthenticated RCE on Windows-hosted servers. Affects 10.0.0 → 15.5.23 and <16.3.2.
[v4.13.5](https://github.com/honojs/hono/releases) (Aug 26) fixes GHSA-crvj-82cr-hjcx (query parser reads params *after* the URL fragment → cache-key and proxy interpretation differentials), GHSA-g6gw-c38x-mqfc (unbounded dot-notation nesting in `parseBody()` → memory exhaustion), and GHSA-gqvv-2mrq-wpjv (incomplete fix for CVE-2026-39408 path traversal in `toSSG()`).
TS [v0.121.0](https://github.com/anthropics/anthropic-sdk-typescript/releases) (Aug 26) adds Standard Schema support for structured outputs and tools, Organization API endpoints, and an `updates` thinking display mode; v0.122.0 (Aug 27) moves beta files/skills namespaces to GA shapes and drops dated beta header pins. Python mirrors this in [v1.1.0/v1.2.0](https://github.com/anthropics/anthropic-sdk-python/releases) (Aug 26–27), on top of the v1.0.0 httpx2 breaking migration (Aug 20).
[Cloudflare changelog](https://developers.cloudflare.com/changelog/) — Aug 25: new `retryAlarm: false` option suppresses automatic alarm retries when `ctx.abort()` kills an alarm. Aug 28: Dynamic Workers concurrency per Durable Object raised from four to ten.
[v0.33.0](https://github.com/ollama/ollama/releases) (Aug 21) adds Claude Desktop integration, letting Ollama act as a third-party gateway provider; v0.33.1 (Aug 26) adds MLX support for Qwen3.8 Flash Next plus structured output; v0.33.2 (Aug 27) fixes the Claude Desktop proxy dropping during catalog updates.
[laude.org](https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents) — Laude Institute + MIT, Aug 24 2026; [HN 123pts/55 comments](https://news.ycombinator.com/from?site=laude.org). <10K lines, agents that think continuously instead of dormant-between-turns. Memory = append-only JSONL trajectory + a compaction algorithm with *exponentially decaying resolution* (recent thoughts verbatim, older progressively summarized) over a DAG so agents can re-enter memory at varying granularity. Test agent diagnosed and fixed its own broken recall in 48 min. Runs $1–2/hr.
[lighthousenewsletter.com](https://www.lighthousenewsletter.com/p/rag-is-simpler-than-you-think) — hit HN Aug 26 2026 at **509 pts / 216 comments** (post itself is June-dated; the debate is this week). Claims 60% of systems should be full-text + LLM query rewriting, rewriting costs ~$0.001/query, hybrid search adds 200–500ms, and agentic query decomposition was "15x cheaper, better quality." Core attack: embeddings force chunking decisions and re-indexing on model deprecation.
[blog.modelcontextprotocol.io](https://blog.modelcontextprotocol.io/posts/mcp-roadmap/) — Aug 22 2026. Next-cycle priorities: agentic messaging primitives (long-running loops, mid-flight steering), extending stateless HTTP to stdio/local servers, **agent identity via DPoP + Workload Identity Federation**, and **progressive tool discovery** to manage tool-count scale. Follows the 2026-07-28 spec that deleted the initialize handshake and Mcp-Session-Id, and deprecated Roots/Sampling/Logging.
[vercel.com/blog/introducing-run](https://vercel.com/blog/introducing-run) — Aug 25 2026. Agent-written TS executes in a fresh QuickJS context in a worker thread, no Node/network access; capabilities exposed only via serialized `hostFunctions`; execution can pause for human approval and resume without redoing work. Shipped alongside "The end of credential sprawl for agents" (same day).
["Small Models Have Arrived"](https://calv.info/small-models-have-arrived), Calvin French-Owen, Aug 26 — **792 pts / 345 comments**; a personalized-aggregation eval fell from ~$1 to ~$0.10, email search across thousands of messages costs "tens of cents," ~100 tps. Same week: [GLM-5.3 open-weight on HF](https://huggingface.co/zai-org/GLM-5.3) (793 pts) and [GLM-5.3-Flash](https://z.ai/blog/glm-5.3-flash) (1128 pts / 575 comments).