[volcengine/OpenViking](https://github.com/volcengine/OpenViking) hit 34,039 stars (+3,078 in 7d). Stores memory/resources/skills as a virtual filesystem under a `viking://` protocol — agents run `ls`/`tree`/`find` instead of querying an opaque vector index. Three-tier loading (L0 ~100-token abstract → L1 ~2k overview → L2 full) with observable retrieval trajectories. Backed by a VLDB 2026 paper; AGPLv3 core, Apache-2.0 CLI. Ollama supported.
[anthropics/claude-plugins-community](https://github.com/anthropics/claude-plugins-community) went 0→2,479 stars this week. Read-only mirror of the community marketplace for Claude Code *and Claude Cowork*, synced nightly from Anthropic's internal review pipeline; submissions go through [clau.de/plugin-directory-submission](https://clau.de/plugin-directory-submission) and pass automated security scanning. Direct PRs are auto-closed. MCP servers are a separate registry.
[akitaonrails/ai-memory](https://github.com/akitaonrails/ai-memory) at ~5k stars. Rust MCP server (stdio + Unix socket) over a git-versioned Markdown wiki with SQLite/FTS5, entity-match and graph-neighbor RRF fusion. Exposes `memory_query`, `memory_write_page`, `memory_handoff_accept`; 15+ agent platforms. MIT.
[apache/maka](https://github.com/apache/maka) — TypeScript, Node 22+, SQLite, Apache-2.0, now in ASF incubation. Append-only durable log of model interactions and tool calls; old tool output is dropped from prompts *without* deleting the evidence. Ships Desktop (Electron+React), TUI/CLI, and an eval harness. macOS Apple Silicon early build; no official release yet.
[tinyhumansai/openhuman](https://github.com/tinyhumansai/openhuman), 38.6k stars, Rust, GPL-3.0. "Memory Tree": scored Markdown compressed into SQLite, mirrored as an editable Obsidian vault. Privacy Mode forces local-only inference; points at Ollama or your own key. Claims 5,000+ MCP servers and 100+ OAuth integrations.
Released Aug 25, 2026. Fixes GHSA-p293-qw3h-jr36 (unauth RCE on Windows-hosted servers) and GHSA-2xp9-vwfh-vxw4 (unauth RCE in the Image Optimization API when AVIF files are used). Canaries after this briefly disabled AVIF entirely, then re-enabled it — a sign the image pipeline fix was invasive.
Aug 26, 2026. GHSA-crvj-82cr-hjcx: query parser reads params *after* the URL fragment, creating cache-key and proxy interpretation differentials (cache poisoning / WAF bypass class). GHSA-g6gw-c38x-mqfc: unbounded dot-notation nesting in `parseBody()` → heap exhaustion from small bodies. GHSA-gqvv-2mrq-wpjv: incomplete fix for CVE-2026-39408, `toSSG()` still writes outside the output dir.
v0.121.0 (Aug 26) adds **Standard Schema support for structured outputs and tools** (use Zod/Valibot/ArkType directly as tool schemas), an `updates` thinking display mode (beta), and Organization API endpoints. v0.122.0 (Aug 27) moves beta files/skills namespaces to GA shapes and drops dated beta header pins; `webhook.unwrap()` now *requires* headers — a breaking change.
4.127.0 (Aug 27) adds `workflows[].concurrency.limit`; 4.126.0 (Aug 25) adds Workflow default retention config for instance persistence, plus PR metadata on preview deployments (GitHub Actions/GitLab/CircleCI) and an experimental `wrangler hyperdrive planetscale signature`. Preview deploys switched to multipart form data for large sourcemaps.
v0.33.0 (Aug 21) makes prefill restore points "trustworthy by construction" so cancelled/retried prefills resume instead of reprocessing from zero, fixes KV-cache invalidation on recurrent-layer models, and adds Claude Desktop as a third-party gateway provider. v0.33.1 (Aug 26) adds Qwen 3.8 Flash Next and MLX structured outputs. v0.33.2 (Aug 27) is pre-release.
[level1techs, Aug 16–27](https://forum.level1techs.com/t/why-your-local-llm-feels-dumber-than-it-is/253917) — 507 pts / 205 comments on HN Aug 22. Teacher-forced decoding across H200/B200/RTX 6000 and vLLM/TensorRT-LLM: INT4 KV-cache quantization *breaks tool calls* while BF16 holds; NVFP4 hits ~50% token disagreement at 88k context; W4A16 fails tool calls outright; W8A16 wins. Attention FP grouping differs per GPU, so logits diverge on identical weights.
[lighthousenewsletter.com](https://www.lighthousenewsletter.com/p/rag-is-simpler-than-you-think) — 494 pts / 206 comments, hit HN Aug 26 (post itself dated June 2026; the debate is this week). Claim: ~60% of systems should stop at full-text + LLM query rewriting; chunking strategy is self-inflicted complexity; reranking 50 docs costs 200–500ms.
[earendil.com, Aug 20](https://earendil.com/posts/what-is-a-harness/) — 578 pts / 181 comments, HN Aug 23. Decomposes an agent into system prompt, tools, agentic loop, and a *translation layer* for provider portability; argues harness ownership shifts power from labs to users.
Calvin French-Owen, [Small Models Have Arrived](https://calv.info/small-models-have-arrived) (Aug 26/27, 649 pts) — complex personalized tasks at ~$0.10 vs ~$1 for Sonnet-class. Same week: [FT: Anthropic's best model struggles to attract users as cheaper tools thrive](https://news.ycombinator.com/item?id=49417884) (816 pts / 701 comments), plus GLM-5.3-Flash (1113 pts) and Qwen3.8-Flash-Next (693 pts) launches.
[vercel.com/blog, Aug 27](https://vercel.com/blog/the-best-workflow-engine-is-a-programming-language) — durable execution from ordinary sequential TypeScript via compiler directives; runs are pinned to deployments; needs only "your app, a database, a queue." Shipped alongside Run SDK (sandboxed agent eval, Aug 25) and Vercel Connect (short-lived scoped agent tokens, Aug 25).