[github.com/volcengine/OpenViking](https://github.com/volcengine/OpenViking) — 32.2k stars, **+3,447 this week**, #1 trending Python. Pitches itself as unifying "Agent Memory, Knowledge RAG and Skills." Notable architecture: context is a virtual filesystem under a `viking://` protocol that agents browse deterministically instead of querying a black-box vector store, with three-tier loading (L0 ~100-token abstract → L1 ~2k overview → L2 full content) claiming 34–91% token reduction. AGPLv3 core, Apache-2.0 CLI.
[github.com/semantica-agi/semantica](https://github.com/semantica-agi/semantica) — 10.4k stars, **+2,450 this week**. Polyglot graph storage (Neo4j, FalkorDB, Apache AGE, Neptune) *plus* swappable vector stores including pgvector, and ships an MCP server (`python -m semantica.mcp_server`) exposing entity extraction, decision recording, and graph analytics. MIT.
[github.com/RyanCodrai/turbovec](https://github.com/RyanCodrai/turbovec) — 16.2k stars, **+1,465 this week**. Rust implementation of Google Research's TurboQuant: 10M docs at 31 GB float32 → 4 GB. Claims 3.4× faster search than FAISS IndexPQFastScan at 4-bit, 4.6–15.1× faster insertion, and O(1) removal (microseconds vs FAISS's milliseconds-to-seconds). MIT, `pip install turbovec`.
[github.com/jundot/omlx](https://github.com/jundot/omlx) — 20.4k stars, **+1,597 this week**. Drop-in replacement for **both OpenAI and Anthropic APIs**, continuous batching via mlx-lm's BatchGenerator, and a two-tier RAM→SSD KV cache (safetensors) that survives server restarts. Apache-2.0. Corroborating heat: "Why your local LLM feels dumber than it is" hit HN at 359 points on 2026-08-22.
Published 2026-08-20. HTTP layer moves to httpx2 (API-compatible fork); requires Python 3.10+; removes legacy Text Completions, and `temperature`/`top_p`/`top_k` on Messages methods; async `.with_raw_response` now needs `await response.parse()`; `AnthropicBedrock` errors instead of defaulting to `us-east-1`. Migration guide with before/after snippets in `MIGRATION.md`.
2026-08-19 changelog. Files API and Agent Skills exit beta (no more `files-api-2025-04-14` / `skills-2025-10-02` headers). New `browser_toolset_20260801` drives a browser your app hosts, reading the accessibility tree rather than screenshots. Managed Agents sessions in self-hosted sandboxes can now attach **memory stores**, mounted at `mount_path` and synced back by the SDK workers.
2026-08-20. `@cloudflare/vitest-pool-workers` → `@cloudflare/vitest-plugin` v1.0.0, with a codemod (`npx @cloudflare/codemods vitest:pool-workers-to-vitest-plugin`) that rewrites package.json, imports, and tsconfig `types`. Wrangler 4.125.0 adds a `connect` trigger delivering raw TCP to a Worker's `connect(socket, env, ctx)` handler — the doc example literally listens on port 5432 — plus container support in Worker previews.
2026-08-20. Bun is now written in Rust ("Claude Code has been using Bun's Rust port for months"); Claude Code's p99 CPU dropped 24% → 10%. Adds `bun test --parallel[=N]` with merged coverage, `Bun.WebView` (headless automation on system WebKit/Chrome, no Puppeteer), `Bun.Image`, 1,517 more passing Node tests, ~48% less memory for HTTP servers.
Tagged 2026-08-21, still prerelease. Fixes a hang when agent clients cancel long prefills; cancelled prefills now keep restore points so retries resume instead of restarting. Notes a case where a request matching 46k of 47k tokens reprocessed from zero on recurrent-layer models. Also disables Claude Code's token-countdown system message that was invalidating the KV cache every request. Adds Claude Desktop integration with per-model toggles.