+4,746 stars this week (8.8k total), hit **#1 on GitHub Trending Aug 10, 2026**. Polyglot backends — Neo4j, Apache AGE, **pgvector**, Qdrant — plus a full MCP server, deterministic reasoning (Datalog/SPARQL/Rete) and W3C PROV-O provenance on every fact. No named enterprise adopters listed.
Released **Aug 10, 2026**; 1,203 HN points / 638 comments. 29.6B dense + 1.8B ViT encoder, 131k context, Apache 2.0, built for tool use + failure recovery on 24GB VRAM. **Ollama 0.32.7 shipped same-day support**, plus llama.cpp, MLX, vLLM.
+742 stars this week (29k total, 313 open PRs). Exposes agent memory/resources/skills as one navigable filesystem via a `viking://` protocol, with L0/L1/L2 tiered detail so agents pay tokens only for the depth they need, and *observable retrieval* (the browse trajectory is preserved and debuggable). AGPLv3 core.
+3,389 stars this week (22.8k total), MIT. Splits memory into four assets (Chat Memory, Skills, Wiki, CodeGraph), layered L0-raw → L3-personas, with "Fixed Binding + ACL" governing which agent sees what (private/team/restricted) and a proxy for zero-code framework integration.
+1,135 stars this week (4.6k total). Tree-sitter → Memgraph + Qdrant, exposed via an MCP server for Claude Code. New this cycle: **runtime call tracing** (runs your test suite, merges actual call edges into the graph to catch dynamic dispatch static analysis misses), Ruby via ast-grep, and AST structural search/replace.
Released Aug 13, 2026 across all supported branches + 19 Beta 3. Fixes 28 security issues. Standouts: [CVE-2026-14666](https://www.postgresql.org/support/security/CVE-2026-14666/) — row-level security keeps using *cached* policies after role membership/ownership changes, so a session can keep reading and writing rows it was just revoked from; [CVE-2026-14676](https://www.postgresql.org/support/security/CVE-2026-14676/) — heap overflow in `pg_stat_statements` giving RCE as the DB OS user via crafted array constants; CVE-2026-14663 — pgcrypto silently encrypting to cleartext for OpenSSL-disabled ciphers.
[anthropic-sdk-typescript v0.117.0](https://github.com/anthropics/anthropic-sdk-typescript/releases) and [anthropic-sdk-python v0.122.0](https://github.com/anthropics/anthropic-sdk-python/releases), both Aug 13, 2026: "add `output_behavior` to dream creation (create a new memory store or update the input store in place)." Dreams take a memory store + 1–100 session transcripts and emit a deduplicated, contradiction-resolved store. Note: the [Dreams doc](https://platform.claude.com/docs/en/managed-agents/dreams) still says "the input store is never modified" — the docs lag the SDK on this. Gated by `dreaming-2026-04-21`.
Aug 13, 2026 — `[email protected]` switches `--framework=next` to scaffold [vinext](https://github.com/cloudflare/vinext) (a Vite plugin reimplementing the Next.js API surface) by default; OpenNext demoted to opt-in `--variant=opennext` for projects needing standard `next build` output.
Aug 13, 2026. Mock Access identities in `wrangler.json` so `ctx.access.getIdentity()` works under `wrangler dev`. Paired with a changelog entry adding worker-level and account-level Access policies covering all domains automatically. Also renames `--ignore-defaults` → `--ignore-base-config`, and adds monthly prompts to update Cloudflare agent skills.
Six releases Aug 11–15. v0.32.12 adds Qwen 3.8 27B with Apple Silicon optimization (agentic/long-horizon gains); v0.32.9 adds NVIDIA Nemotron 3.5 Lightning, a 30B MoE tuned for agent workloads; v0.32.11 adds web search to the OpenAI-compatible API. Cloudflare added Qwen 3.8 27B (262K ctx) to Workers AI the same week.
[Cloudflare, Aug 14](https://blog.cloudflare.com/mcp-security-updates/) — Gateway classifies MCP via `MCP-Protocol-Version`, plus new `Mcp-Method`/`Mcp-Name` headers on *every* request in spec rev 2026-07-28 (not URL heuristics). Ships `experimental.is_mcp` as a policy selector, a shadow-MCP-server dashboard, and Agents SDK v0.20.0 with stateless-protocol support + auto negotiation.
[deepseek.com/harness](https://deepseek.com/harness/en/), Aug 13 — 737 pts / 309 comments. "Agent = Model + Harness." Everything is a swappable plugin: models, tools, sessions, sandboxes, **storage**, loops, subagent scheduling. Append-only event stream records every context injection; "Code mode" lets the model emit one TypeScript program instead of N tool calls. HN pushback: plugin-ecosystem rot ("nightmare of incompatible, deprecated plugins") and TS/Node choice.
[Simon Willison, Aug 16](https://simonwillison.net/2026/Aug/16/qwen-38-27b/) (775 pts) on the [Apache-2 release](https://huggingface.co/Qwen/Qwen3.8-27B-FP8) (1425 pts). 17GB Q4_K_M, reliable tool-calling, drove a coding agent end-to-end. But it ships defaulting to `xhigh`: 22,276 reasoning tokens / 21 min for one SVG; reasoning off → 137s. `--spec-type draft-mtp` in llama.cpp gave ~72% speedup. He tested LM Studio and llama-server, not Ollama.
[TechCrunch, Aug 16](https://techcrunch.com/2026/08/16/stripe-will-reportedly-acquire-ai-gateway-startup-openrouter-for-7b/) — 462 pts / 290 comments. 5.4× its $1.3B May valuation. HN split: "routing is the expertise" vs. "changing a URL and API token is minimal effort." One theory — defensive, since OpenAI+OpenRouter volume is ~5% of Stripe's TPV.
[platform.claude.com release notes](https://platform.claude.com/docs/en/release-notes/system-prompts), Aug 16 — 746 pts / 280 comments. Top argument: "the longer the system prompt and the more you cram in it, the worse the model does." Also: "instructions containing contradictions lead to diminished quality *outside* the scope of the contradiction." Recommended method — start empty, tune only where it errs.