[github.com/volcengine/OpenViking](https://github.com/volcengine/OpenViking) — +2,444 stars this week, 31.4k total, #2 on weekly Python trending. ByteDance/Volcengine. Stores memory + resources + skills under a `viking://` protocol so agents `ls`/`tree`/`find` their context instead of querying a vector store. Three-tier loading (L0 ~100-token abstract → L1 ~2k overview → L2 full). MCP server + HTTP + Python/Node/Rust SDKs; AGPLv3 core.
[github.com/akitaonrails/ai-memory](https://github.com/akitaonrails/ai-memory) — +1,952 stars this week, 3.8k total, #4 weekly Rust. SQLite + FTS5 + git-versioned markdown; explicitly "no vector database to babysit." MCP server exposing `memory_query` / `memory_write_page` / `memory_feedback`; supports handoff across Claude Code, Codex, Cursor, Gemini CLI, OpenCode, +10 more. MIT.
[github.com/semantica-agi/semantica](https://github.com/semantica-agi/semantica) — +3,674 stars this week to ~10k; was 3,435 on Aug 10, so it roughly tripled in ~11 days. Polyglot graph backends (Neo4j, FalkorDB, Apache AGE) + vector backends incl. **pgvector**, full MCP server, W3C PROV-O provenance per fact, deterministic reasoning (Datalog/SPARQL, no LLM). MIT.
[github.com/RyanCodrai/turbovec](https://github.com/RyanCodrai/turbovec) — +1,110 stars this week, 16.1k total. Implements Google Research's TurboQuant: no training phase, 1536-dim float32 → 384 bytes at 2-bit, 10M docs from 31 GB → 4 GB RAM. Claims 3.4× faster than FAISS IndexPQFastScan at 4-bit with comparable R@1. MIT. Note: README benchmarks **do not** include pgvector or HNSW.
[github.com/jundot/omlx](https://github.com/jundot/omlx) — +1,388 stars this week, 20.1k total. Native Swift menubar server: continuous batching, tiered hot-RAM/cold-SSD KV cache, LLMs + VLMs + embeddings + rerankers, tool calling and structured output. Drop-in **OpenAI and Anthropic** API compatibility. Apache 2.0.
[Aug 19 release notes](https://platform.claude.com/docs/en/release-notes/api) — Managed Agents sessions in a **self-hosted sandbox** can now attach memory stores; the Python/TS/Go SDK workers download each store to its `mount_path` and sync agent writes back. Same day: Files API and Skills API (`/v1/skills`) dropped their beta headers (GA), `computer_toolset_20260801` GA, new `browser_toolset_20260801`, and `allowed_domains`/`blocked_domains` on agent `web_search`/`web_fetch`. Shipped in `anthropic-sdk-typescript` v0.119.0/0.120.0 and `anthropic-sdk-python` 0.124.0/0.125.0.
[Aug 20](https://github.com/cloudflare/workers-sdk/releases/tag/wrangler%404.125.0) — new `connect` trigger lets a Worker accept raw TCP (`"connect": [{"protocol":"tcp","port":5432}]` → `connect(socket, env, ctx)`, experimental flag). Also container support in Worker previews, `WorkflowInstance.delete()`/`deleteBatch()` (100/op), and a fix for `kv bulk put` corrupting `base64: true` binary values under `wrangler dev` (12-byte PNG header read back as 20 bytes). `@cloudflare/vitest-pool-workers` renamed to `@cloudflare/vitest-plugin` v1.0.0 with a codemod.
[Aug 20](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v1.0.0) — first 1.0; `⚠ BREAKING: client: upgrade to httpx2`, see MIGRATION.md. Also exports `ServiceUnavailableError`/`DeadlineExceededError` at root, and fixes sync session tools blocking the worker thread (0.123.0).
[Aug 13](https://www.postgresql.org/about/news/postgresql-186-1711-1615-1519-1424-and-19-beta-3-released-3365/) — 28 security fixes and 110+ bugs, including CVE-2026-19385, a `pg_dump` heap buffer overflow enabling arbitrary code execution (CVSS 8.8, affects 14/15/16). 18.5 was skipped due to a regression.
[Aug 20](https://bun.com/blog/bun-v1.4) — +1,517 Node test-suite tests passing (`node:http`, `node:fs`, `node:sqlite` at 97–100%); Next.js 16, Playwright, vitest and OpenTelemetry now run on Bun. Zig→Rust rewrite: 5× lower idle CPU, 13–48% less memory on HTTP apps, 2–2.5× faster startup. New `bun test --parallel`/`--shard`/`--changed`, `Bun.cron()`, `Bun.WebView`.
[HN 432pts/261 comments, Aug 19](https://news.ycombinator.com/item?id=49361279). Commenters put hard numbers on where "just use Postgres" breaks: HNSW/IVFFlat degrade past ~100M vectors, and adding pgvector to a busy OLTP instance "will either trash your caches, or take over your CPU so that workloads that used to work fine stall." PG core contributor **anarazel** pushed back on the genre itself.
[HN 1437pts/792 comments, Aug 14](https://news.ycombinator.com/item?id=49299605); [Simon Willison's writeup, Aug 16](https://simonwillison.net/2026/Aug/16/qwen-38-27b/) (795pts). 17GB Q4_K_M fits a 128GB M5 Max; drives the Pi coding agent with reliable tool-calling. Two caveats: 15–30 tok/s (memory-bandwidth bound), and it defaults to `xhigh` reasoning — Simon says start at low/none.
[Cloudflare, Aug 14](https://blog.cloudflare.com/mcp-security-updates/) — Gateway now classifies MCP by the `MCP-Protocol-Version` header (`experimental.is_mcp == true`), targeting two named failure modes: **Shadow MCP** and **Portal Bypass**. Same day: [Access for Workers](https://blog.cloudflare.com/workers-protected-by-access/).
[OpenRouter announcement, Aug 19, HN 949pts/486](https://news.ycombinator.com/item?id=49364559) + [TechCrunch report, Aug 16, 485pts](https://news.ycombinator.com/item?id=49323381). Debate splits on dependency risk: `johnbarron` cites 50ms+ latency, shared capacity pools, data visibility; `ApolloFortyNine` questions ~15% markup vs direct provider access. Defenders cite unified keys with budget caps across providers.
[Anthropic, Aug 14](https://www.anthropic.com/news/claude-text-watermark); [Gruber's response, Aug 16, HN 822pts/728](https://news.ycombinator.com/item?id=49324087) calling it "a perversion of writing." Separately Anthropic [published Claude's system prompts](https://platform.claude.com/docs/en/release-notes/system-prompts) (758pts, Aug 16).