[Show HN, 180 pts / 49 comments, posted ~26 Aug](https://news.ycombinator.com/item?id=49437049). Repo at 518 stars, written in Zig, MIT, zero deps, already at v0.11.1 with Python **and** TypeScript/Node bindings. Claims 0.83 ms vector search at 1M vectors with 100% recall, 0.13 μs node lookups. Comment thread compares it directly to Neo4j, Kuzu, SurrealDB; author was candid about benchmark errors and shipped two releases in 24h.
+3,691 stars this week (33.7k total), #11 on weekly trending overall. Stores agent memory as a virtual filesystem under a `viking://` protocol with three tiers (L0 abstract → L2 detail), loaded on demand. Retrieval is directory-recursive: vector search finds a directory, then drills down, preserving the browse trajectory for debugging. Ships MCP client integration and Ollama support; AGPLv3 with a Volcano Engine managed SaaS.
+1,714 stars this week (4.9k total). Pitch: "quit Claude Code mid-task, start Codex in the same directory, continue without re-explaining." Layered retrieval — SQLite FTS5 + entity-match ranking + graph-neighbor expansion + *optional* embeddings — with authority-aware weighting that favors canonical decision pages over episodic session logs. Git-versioned markdown wiki is the source of truth; runs as an MCP server across Claude Code, Codex, Devin CLI.
+1,759 of its 2,336 stars arrived this week — effectively a brand-new repo. Read-only mirror of the reviewed marketplace for Claude Code and Cowork; PRs are auto-closed, submissions go through [clau.de/plugin-directory-submission](https://clau.de/plugin-directory-submission) with automated security scanning, manual review, then nightly sync. Install via `claude plugin marketplace add anthropics/claude-plugins-community`.
+1,818 stars this week (38.4k total); was GitHub's #1 trending repo for nine consecutive days after launch. Rust backend with TypeScript components, GPL-3.0. Memory Tree compresses user data into scored Markdown in SQLite, mirrored as an Obsidian vault for transparency. Privacy Mode forces on-device processing, secrets in OS keyrings; runs Ollama, BYO keys, or hosted. Checkpointed graph runs let agents pause and resume.
Released Aug 25, 2026. `GHSA-2xp9-vwfh-vxw4` — unauthenticated RCE in the Image Optimization API when AVIF files are used (Critical); `GHSA-p293-qw3h-jr36` — unauthenticated RCE on Windows-hosted servers (Critical). Both reported by eps1lon, both patched in the same pair of releases. The advisory pages did not render CVE IDs or exact affected-version ranges, so pin from the release notes.
TS `v0.121.0` (Aug 26) adds Standard Schema support for structured outputs and tools, an `updates` thinking display mode (beta), Organization API endpoints, and fixes the tool runner to continue on `pause_turn`. Python `v1.0.0` (Aug 20) upgrades the client to **httpx2** with breaking changes (see MIGRATION.md); `v1.1.0` (Aug 26) follows with the same feature set. Preceding these, `v0.119.0`/`v0.124.0` (Aug 19) moved the **Files and Skills APIs to GA**.
Cloudflare changelog, Aug 22, 2026 — Wrangler and the Cloudflare API MCP server support optional OAuth scopes, letting users "choose which specific permissions to grant" rather than approving all requested scopes at authorization time.
`4.126.0` (Aug 25) adds Workflows `default_retention` for instance retention periods, FUSE support for local Durable Object containers, PR metadata on preview deployments, and an experimental `wrangler hyperdrive planetscale signature`. `4.127.0` (Aug 27) adds `workflows[].concurrency.limit` and switches `wrangler preview` to multipart module upload, shrinking payloads for large sourcemaps.
`0.33.0` (Aug 21) makes "prefill restore points trustworthy by construction," fixes KV-cache bugs affecting token-countdown systems, and handles cancelled long prefills. `0.33.1` (Aug 26) adds Qwen 3.8 Flash Next support and avoids Metal GPU timeouts loading models from slow storage.
[MCP roadmap, Aug 22](https://blog.modelcontextprotocol.io/posts/mcp-roadmap/) — 269 pts / 161 comments. Five priorities: agentic messaging primitives (Tasks, subscriptions, mid-flight steering), HTTP-native transport everywhere (Streamable HTTP even over stdio), agent identity via **DPoP + Workload Identity Federation** (killing long-lived API keys), and **progressive tool discovery**. The top HN thread pushes back hard: `skinfaxi` and `wilj` both say they've abandoned lazy-loading MCP tools for Cloudflare's code-mode. Pairs with [Cloudflare's task-based OAuth consent](https://blog.cloudflare.com/task-based-oauth-consent/) (Aug 20), which names MCP servers as the motivating case for optional/declinable scopes.
[Why your local LLM feels dumber than it is](https://forum.level1techs.com/t/why-your-local-llm-feels-dumber-than-it-is/253917), Aug 22 — 507 pts / 207 comments. Switching vLLM's attention kernel (FA2/FI/TRT) on identical weights reproducibly broke tool calls. NVFP4 hit ~50% token disagreement by 88k context; INT8 W8A16 stayed faithful. INT4 KV-cache flips top tokens mid-tool-call.
[GLM-5.3-Flash](https://news.ycombinator.com/item?id=49449507), Aug 26 — 1036 pts / 523 comments; thread pins pricing at **$0.15 in / $0.50 out / $0.03 cached** per 1M. Same week: [Qwen3.8-Flash-Next](https://news.ycombinator.com/item?id=49448210) (671 pts) and [FT: "Anthropic's best AI model struggles to attract users as cheaper tools thrive"](https://news.ycombinator.com/item?id=49411102) (816 pts / 700 comments).
[earendil.com, Aug 20](https://earendil.com/posts/what-is-a-harness/) — 576 pts / 179 comments. Decomposes a harness into system prompt, tools, agentic loop, translation layer; argues the model layer commoditizes and the harness is where user-owned value accrues. `theturtletalks`: "If LLMs are electricity, harnesses are the electronics."
[RAG Is Simpler Than You Think](https://www.lighthousenewsletter.com/p/rag-is-simpler-than-you-think) — 461 pts / 194 comments on Aug 26 (note: the **post itself is dated June 10**; only the discussion is fresh). Argues BM25 + LLM query rewriting beats embeddings for most corpora, cites 200–500ms embed latency and model-deprecation risk. Thread splits: `7734128` counters that 90% of doc RAG *should* be embeddings-first; `pantsforbirds` reports keyword search wins on technical content.