Quiet news day, but two of your own forwards point at the same thing the bigger feeds are tracking: the Claude Code roadmap is shifting under your feet (Mythos 1 model + Memory Files rework both staged this week) and the home-lab move off Docker Desktop is becoming a mainstream conversation, which lines up with your post-hardware-install plan for Apex.
One smaller item but worth flagging: Voxtral 2 + Parakeet V3 benchmarks landed publicly enough that they're now in the casual "best 2026 STT" comparison posts, which means you'll see them in random Reddit/HN threads. Your Bidet Quick swap-test backlog just got more data.
Why it matters to you: The model string claude-mythos-1-preview briefly appeared as a toggle in the public Claude Code interface before being pulled. Mythos has been the locked-down Project Glasswing cyber-defense model since April; it moving toward Claude Code means the agent rotation you use (Yaya / Cursor / Codex) gets a new heavyweight option soon. Worth watching for the model picker to grow another row.
Why it matters to you: Author moved off Docker Desktop to native Proxmox LXCs and cites the exact failure modes you hit on Apex (mount weirdness, RAM overhead, nested-container security holes). PVE-Helper Scripts do one-liner installs of self-hosted apps. Not an urgent migration, but a real option to keep in mind when you re-stand-up TP3 services after the hardware install, particularly if Docker Desktop keeps eating bind mounts.
Why it matters to you: Anthropic is testing a dual-mode memory: classic single-summary vs the new "Memory Files" - a built-in personal wiki of topic-scoped documents the assistant consults selectively. That is precisely the architecture your MEMORY.md index already uses (one-liner pointers to per-topic detail files). Two reads: (1) validates the pattern, (2) once it ships, Claude.ai web UI gets the same shape G16's local memory has - probably means you can stop maintaining the manual scaffold for the web side.
Why it matters to you: Scorecard that tracks a user's AI capabilities inside Claude itself - aimed at "fluency" assessment. Direct overlap with the teacher-facing curriculum lane you've been mapping (St. Francis). If they ship it consumer-side first, it's a reference design for what an AI-literacy unit could measure. Worth a 5-minute look when it lands; do not extrapolate into spec.
Why it matters to you: Mainstream "best open STT 2026" roundups now cite Voxtral 2 + Parakeet V3 + Moonshine as the trio above Whisper. Voxtral is 4B Apache 2.0 with native real-time streaming but only 13 languages. Parakeet V3 wins long-form. Moonshine is 245M and Apple-Silicon-tuned for low-latency voice agents. All three are already in your Bidet Quick swap-test backlog from the May 22 voice radar - now there's enough public benchmarking to make the A/B less speculative.
Why it matters to you: You already use Cursor in the agent rotation for notecard Pass 3/4. Composer 2.5 is the in-house coding model getting RL-corrected for behavioral calibration. Mostly relevant as a "your Cursor will feel different this week, don't blame Yaya" heads-up.