[volcengine/OpenViking](https://github.com/volcengine/OpenViking) — +4,211 stars this week (~33.5k total), #1 on GitHub Trending/Python. ByteDance-backed. Unifies memory + knowledge RAG + skills behind a `viking://` filesystem the agent browses with `ls`/`tree`/`find` instead of opaque vector queries, with L0/L1/L2 tiered loading (summary → ~2k-token overview → full detail). Sessions auto-distill into long-term memory post-commit. AGPLv3 core, Ollama embeddings supported.
[semantica-agi/semantica](https://github.com/semantica-agi/semantica) — +1,898 stars this week (~10.9k). Python, MIT. Pitches itself as "open source Palantir for AI agents": context graphs, decisions as first-class objects with causal links, W3C provenance. Backends include **Neo4j and pgvector** directly; ships a `semantica-mcp` server exposing extraction/reasoning/graph-query tools. No LLM needed for graph construction.
[Show HN: LatticeDB – Like SQLite but for graph databases](https://news.ycombinator.com/item?id=49437049) — 157 pts / 41 comments, front page Aug 25. Zig, zero deps, MIT, ~387 stars. Single Cypher query spans graph traversal, vector similarity, and full-text. Claims 0.83ms mean @1M with 100% recall@10, 70x SQLite on variable-path traversal. **TypeScript/Node, Python, and Go bindings.** Author explicitly names agentic memory and graph RAG as targets; concurrency file-locking is still missing.
[akitaonrails/ai-memory](https://github.com/akitaonrails/ai-memory) — +2,073 stars this week, #3 Rust trending. Pitch: quit Claude Code mid-task, resume in Codex in the same dir without re-explaining. Markdown wiki in **git** as source of truth, SQLite+FTS5 for index, embeddings alongside — explicitly "no vector DB to babysit." Ships as an MCP server with a support matrix spanning Claude Code, Codex, Cursor, Gemini CLI, Kiro, OpenCode and ~10 more.
[tinyhumansai/openhuman](https://github.com/tinyhumansai/openhuman) — +1,345 stars this week, top-trending within a week of launch. Rust/Node/Tauri. Memory Tree + Obsidian wiki compressed into local SQLite, 20-min auto-sync, 100+ OAuth integrations, MCP servers, Ollama-only Privacy Mode with a single "no inference leaves your machine" switch.
Released Aug 25, 2026. GHSA-2xp9-vwfh-vxw4 (CVSSv4 **9.5**) is unauthenticated RCE in the Image Optimization API via AVIF, rooted in `libheif` under `sharp`; affects **>=10.0.0 <15.5.24** and **<16.3.3**. Vercel disabled AVIF optimization as a stopgap. GHSA-p293-qw3h-jr36 is a second critical RCE on Windows-hosted servers. https://github.com/vercel/next.js/releases/tag/v16.3.3
Aug 20, 2026. HTTP layer moves to `httpx2`; requires Python 3.10+; removes legacy Text Completions, **removes `temperature`/`top_p`/`top_k` from Messages methods**, removes tool-runner `compaction_control`. Async `.with_raw_response` now needs `await response.parse()`; `AnthropicBedrock` errors instead of defaulting to `us-east-1`. Call `httpx2.alias_httpx()` if anything patches httpx (tracing/mocking). https://platform.claude.com/docs/en/release-notes/api
Aug 19, 2026. Files API and `/v1/skills` out of beta — no more `files-api-2025-04-14` / `skills-2025-10-02` headers; headerless requests get the new shape (`expires_in_seconds`/`expires_at`, `page`/`next_page` + `ids[]` filter). Computer use GA as `computer_toolset_20260801` (batch actions, zoom default). New **`browser_toolset_20260801`** drives a browser your app hosts, reading the accessibility tree. Mirrored in TS SDK v0.119.0. https://platform.claude.com/docs/en/release-notes/api
Aug 26, 2026 (today). Query parser didn't stop at the URL fragment → cache-key/proxy interpretation mismatch; SSG path traversal (consecutive `..` segments in `ssgParams` not fully collapsed — prior CVE fix was incomplete); `parseBody({dot:true})` unbounded nesting → memory exhaustion. Maintainers urge immediate upgrade for Cache Middleware / SSG / dot-notation users. https://github.com/honojs/hono/releases
Aug 25, 2026. Workflow bindings gain `default_retention` (separate success/error windows, e.g. `"3 days"`/`"7 days"`); preview deployments now attach PR number/URL/title as annotations; experimental `wrangler hyperdrive planetscale signature`; local DO containers get Docker FUSE. Companion `@cloudflare/[email protected]` consolidates dev bindings under a `dev` object. https://github.com/cloudflare/workers-sdk/releases
[blog.modelcontextprotocol.io/posts/mcp-roadmap](https://blog.modelcontextprotocol.io/posts/mcp-roadmap/) (Aug 22, 269 pts / 160 comments on HN). Protocol-level sessions and the init handshake are being removed so servers scale horizontally without holding state; server-initiated requests get replaced by Multi Round-Trip Requests; `server/discover` + `.well-known` metadata for capability negotiation; progressive tool disclosure instead of dumping the whole catalog; DPoP and Workload Identity Federation replacing long-lived API keys. The 2026-07-28 release already shipped breaking changes.
[Level1Techs forum, Aug 22](https://forum.level1techs.com/t/why-your-local-llm-feels-dumber-than-it-is/253917) (506 pts / 206 comments). Same weights, different CUDA attention kernels → ~10% token disagreement at 88k context. INT4 KV cache caused reproducible tool-call failures (model couldn't close function calls); INT8 recovered. Weight formats compared: INT8 W8A16 best at 0.7% token flips, NVIDIA NVFP4 worst at ~50% by 88k. FlashAttention 2 even picks different group counts per GPU (54 on H200, 62 on B200), changing rounding.
[openrouter.ai/blog, Aug 19](https://openrouter.ai/blog/announcements/openrouter-is-joining-stripe/) (962 pts / 498 comments). 10T+ tokens/day, 400+ models, 10M+ developers. Claims nothing changes for existing integrations and routing stays neutral "regardless of parent company." Closing expected within weeks.
[FT via HN, Aug 23](https://news.ycombinator.com/item?id=49411102) (813 pts / 698 comments); [Simon Willison's note](https://simonwillison.net/2026/Aug/23/anthropics-best-ai-model-struggles-to-attract-users-as-cheaper-t/). Per FT, Anthropic's top-tier model took only ~11% of customer spend since June while a cheaper tier climbed past 50% by late July. Cause given: enterprises adopting model routing — cheap models for simple tasks, expensive ones reserved for hard ones. Separately, [OpenAI cut GPT-5.6 Sol pricing Aug 24](https://developers.openai.com/api/docs/pricing) (335 pts). Could not fetch the FT article itself (paywalled).