+4,005 stars this week (9,701 total), #2 on GitHub Trending overall, pushed 2026-08-20. MIT-licensed, and the backend matrix is almost a description of Sri's stack: Neo4j *and* pgvector both first-class (`semantica[graph-neo4j]`, `semantica[vectorstore-pgvector]`), plus a full MCP server and REST API on :8000. Sells on provenance/auditability, not just recall.
Cloudflare shipped MCP v2 support (blog dated 2026-08-06 — 14 days, just outside the window, but the migration is live now). `Mcp-Session-Id` and protocol sessions are gone; MCP is fully stateless HTTP. New `Mcp-Method`/`Mcp-Name` headers. Dynamic Client Registration deprecated in favor of Client ID Metadata Documents. Server authors must move `McpAgent` → `createMcpHandler`.
+1,659 this week, 30,711 stars, pushed 2026-08-20. Exposes memory/skills via a `viking://` filesystem agents navigate with `ls`/`tree`/`grep` instead of querying a vector store. Three-tier loading: L0 ~100-token summaries → L1 ~2k overviews → L2 full content. **AGPL-3.0** core.
+2,805 this week, 23,377 stars, created 2026-04-07 (fast ramp), pushed 2026-08-15, v2.0.0. Layered model: L0 raw conversation → L1 atomic facts → L2 project scenario → L3 persona, plus a CodeGraph asset. Positions on *team-level* memory and governance.
Aug 18–19, 2026. Files API GA (1 TB/org, configurable `expires_in_seconds`), Skills API GA (`/v1/skills`, no beta header), and **workspace-scoped memory stores** for Managed Agents — memories as files, 2,000 per store, 100 kB each, every write creating an immutable version with point-in-time recovery. SDKs shipped same-day: TS `v0.120.0`, Python `v0.125.0` ("managed agents web search config and self hosted sandbox memory"). Self-hosted sandbox workers mount an attached store at `mount_path` and sync back.
Aug 13, 2026. 28 CVEs incl. `CVE-2026-6464` psql COPY FROM STDIN command injection (8.1), `CVE-2026-14664` regexp heap overflow (8.8), `CVE-2026-15741` expression-deparse SQL injection (8.8). 110+ bug fixes. **Post-update actions required:** `ANALYZE` tables with GIN indexes (parallel builds wrote bad `reltuples`), `REINDEX` btree_gist indexes on float/bit columns. PG14 EOL Nov 12, 2026. PG19 Beta 3 also out.
Aug 13–20. `v0.32.12` Qwen 3.8 27B with an Apple Silicon MLX build (`qwen3.8:27b-mlx`) tuned for repeated coding-agent runs; `v0.32.11` web search in the **OpenAI-compatible Responses API** plus DeepSeek Harness; `v0.32.10` MLX prefill +7–8% and default repeat penalty 1.1→1.0; `v0.32.15` model metadata cache cutting per-request overhead.
Aug 14, 2026. Attach an Access policy to a *Worker* rather than per-domain — covers all its domains and **preview URLs** automatically, or default-on account-wide with explicit public exemptions. Works under `wrangler dev` and via the Workers API.
Aug 18, 2026. New `container` field binds a container bidirectionally to a DO via the exports entry (`containers[].class_name` now optional). Compatibility date now defaults to the shipped workerd's release date instead of today's date, which could exceed workerd's 7-day limit. D1 fixes: CRLF normalization and lowercase `end` in SQL blocks.
Released Aug 16; HN's #3 story of the week (1,432 pts / 792 comments) plus [Simon Willison's writeup](https://simonwillison.net/2026/Aug/16/qwen-38-27b/) (793 pts). Runs at 15–30 tok/s on an M5 Max via llama.cpp/LM Studio, drove Pi coding agents successfully. Simon's caveat: default `xhigh` reasoning effort massively overthinks — set it to `low` for real work.
Aug 14. Gateway now classifies MCP requests via the `MCP-Protocol-Version` header — new `experimental.is_mcp == true` policy selector, an MCP traffic dashboard, and an `mcp_portal` traffic-source selector to distinguish Portal-proxied from direct connections. Free for existing Zero Trust customers. MCP Portals also gained manual OAuth client pre-registration.
Announced Aug 19 (846 pts / 442 comments; [TechCrunch on the leak](https://techcrunch.com/2026/08/16/stripe-will-reportedly-acquire-ai-gateway-startup-openrouter-for-7b/), 481 pts). OpenRouter commits to "same name, same product, same roadmap" and explicit model neutrality — "no single model becomes the default by inertia." Closing in weeks. Related: GPT-5.6 Sol pricing cut 50% on OpenRouter (629 pts).
295 pts / 34 comments, Aug 18. Data-oblivious quantizer, no training phase, online ingest with no rebuilds. 10M docs: 31GB fp32 → 4GB. 3.4–3.5× faster than FAISS IndexPQFastScan at 4-bit; single-vector adds 6–20µs (7–14× faster than FAISS). Python + Rust bindings. No pgvector integration.
Aug 16, Gruber (818 pts / 727 comments). Claude biases word choice toward secret green/red lists for outputs over ~200 tokens, across all models worldwide — per Gruber, **including API output**, not just claude.ai. EU-compliance driven. Objection: it degrades prose and survives copy/paste, so human-edited text can false-flag.