[github.com/akitaonrails/ai-memory](https://github.com/akitaonrails/ai-memory) — +1,030 stars this week (Rust trending), 5.3k total, 1,432 commits. Captures agent lifecycle hooks into a git-versioned markdown wiki, indexed in SQLite FTS5 + optional embeddings, with graph-neighbor RRF reranking. Ships as an MCP server across 25+ agent CLIs. Pitch is vendor handoff: quit Claude Code mid-task, resume in Codex with context intact.
[Show HN, Aug 25 2026](https://news.ycombinator.com/item?id=49437049), 188 points / 54 comments (HN thread itself returned 429 on refetch — points from search index). Zig, single-file property graph with a Cypher subset (MATCH, variable-length paths, MERGE, WITH, UNWIND), native HNSW vector search and BM25 in the same query layer. TS/Python/Go bindings; claimed ~2,819x graph traversal vs SQLite, ~300x vs FTS5.
[github.com/tinyhumansai/openhuman](https://github.com/tinyhumansai/openhuman) — +2,526 stars this week (Rust trending), #1 trending nine consecutive days. "Memory Tree" compresses personal data into scored Markdown in SQLite, syncs to an Obsidian vault, auto-fetches from 100+ services every 20 min. Ollama for fully-local inference with a Privacy Mode switch. GPL-3.0, early beta.
[github.com/abhigyanpatwari/GitNexus](https://github.com/abhigyanpatwari/GitNexus) — +870 stars this week (TS trending). Tree-sitter + LadybugDB build a code knowledge graph; the differentiator is precomputing clusters, execution flows and impact analysis at index time so the LLM gets full context in one tool call instead of multi-hop edge traversal. Runs as CLI+MCP or fully in-browser via WASM.
[github.com/agentgateway/agentgateway](https://github.com/agentgateway/agentgateway) — Rust trending this week (+155 stars, 4.7k total), Linux Foundation–backed, 2,536 commits. Federates MCP tools across stdio/HTTP/SSE/Streamable HTTP, adds JWT/API-key/OAuth, CEL-based RBAC, and OpenTelemetry tracing.
Two Critical advisories published **Aug 25, 2026**: `GHSA-2xp9-vwfh-vxw4` (Unauthenticated RCE in the Image Optimization API when AVIF files are used) and `GHSA-p293-qw3h-jr36` (Unauthenticated RCE on Windows-hosted servers). Patched in v16.3.3 and v15.5.24, both shipped Aug 25. Reported internally by eps1lon. The 16.4 canaries note AVIF was temporarily disabled, then re-enabled after the fix.
Released **Aug 26, 2026**. Changelog: "**helpers:** support Standard Schema for structured outputs and tools" (commit `7880078`). Standard Schema is the cross-library spec that Zod, Valibot and ArkType all implement — so tool params and structured outputs can be defined once from an existing validator instead of hand-written JSON Schema. Same release adds an `updates` thinking display mode (beta) and fixes the tool runner to continue on `pause_turn`.
**Aug 27, 2026** platform release note: in TS SDK 0.122.0 / Python SDK 1.2.0, `client.beta.files` and `client.beta.skills` stop sending the `files-api-2025-04-14` and `skills-2025-10-02` headers and return GA shapes. Behavior changes: `client.beta.skills.delete()` now deletes a Skill *and all its versions*; `BetaSkill` renamed `BetaContainerSkill`. Requests still sending the old headers keep beta shapes.
**Aug 28, 2026** — a Durable Object can now use up to ten Dynamic Workers concurrently, up from four, per DO instance across all concurrent requests. Separately (**Aug 25**), `ctx.abort()` accepts `{ retryAlarm: false }` to suppress alarm retries. Also **Aug 28**: `@cloudflare/[email protected]` fixes tests that construct the same DO many times getting progressively slower and eventually stack-overflowing.
v0.33.1 (**Aug 26**) added Qwen3.8 Flash Next support and **structured output in the MLX runner**; v0.33.2 (**Aug 27**) restored system dark-mode and stopped the Claude Desktop proxy interrupting in-flight requests during catalog updates. v0.33.0 (Aug 21, just outside window) introduced configuring Claude Desktop to use Ollama as a third-party gateway provider.
https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents — published 2026-08-24, HN 123 pts / 55 comments. Authors include Omar Khattab (DSPy), Andy Konwinski, Braden Hancock. Two design ideas matter: a compaction algorithm keeping the *entire* trajectory in context at exponentially decaying resolution (recent verbatim, older progressively summarized), and a trajectory stored as a DAG so the agent retrieves memory at varying granularity. Open source: github.com/laude-institute/headlong.
https://www.lighthousenewsletter.com/p/rag-is-simpler-than-you-think — hit HN front page 2026-08-26 with **512 pts / 216 comments** (post itself dated June 10; the debate is what's fresh). Rafael Pierre argues ~60% of systems should stop at BM25 full-text + query rewriting, that rewriting beats embeddings because you change a system prompt instead of re-embedding a corpus, and that agentic sub-query decomposition ran 15x cheaper than naive retrieval.
https://vercel.com/blog/introducing-run (2026-08-25) — execute untrusted agent-generated JS/TS in isolation, with pauses for human review. Same week: https://vercel.com/blog/the-best-workflow-engine-is-a-programming-language (2026-08-27), durable workflows as ordinary code on existing infra, and Vercel Connect GA replacing stored provider secrets with short-lived scoped tokens.
https://calv.info/small-models-have-arrived — Calvin French-Owen (Segment co-founder), 2026-08-26, HN **796 pts / 348 comments**. Claims a personalized-news task that cost ~$1 on Sonnet-class now runs ~$0.10 on GPT-5.6-Luna at ~100 tok/s; places GLM-5.3 on the Pareto frontier. GLM-5.3 went open-weight 2026-08-28 (HN 800 pts).
https://acceptmarkdown.com/ — HN 176 pts / 108 comments (2026-08-26; site itself timed out, summary from the HN thread). Proposal: content-negotiate Markdown for agents. Cited 360KB HTML → 9KB Markdown. Objections were sharp: bots can't verify the Markdown matches the HTML, prompt-injection surface, and site incentives run the other way.