Give your coding agent a docs 'brain,' not a vector memory
A fast-spreading Oct 3 argument: agents work better rereading Markdown than RAG recall. Plus Codex projectless sessions and new speech models on Vercel.
Copy markdown
Docs your agent rereads beat RAG recall
An Oct 3 post making the rounds argues conversation-memory — RAG over past chats — is stale, context-free, and ranks by similarity rather than truth, so coding agents drift. The fix: a Markdown 'project brain' (an internal/ folder of specs, plans, and an index) the agent reads before work and updates after, turning prompt-build-forget into consult-build-update.
Operator Memory puts the pattern in code
The same author ships aerovato/operator-memory (BSD-3): agents write .operator/ project files and ~/.operator user notes as they work, versioned in git beside your code. It wires into OpenCode and Code Puppy today, with Claude Code and Codex adapters marked pending — a reference build to copy now, not yet a drop-in.
Codex 0.160.0 adds projectless sessions
OpenAI's Codex CLI 0.160.0 lets you start a session anywhere with workspace defaults — no project scaffolding needed — and adds keyboard-navigable history in the agent command center, X11 middle-click paste, and reconnection fixes. Handy for one-off scripts and quick repros.
Microsoft speech lands on Vercel AI Gateway
Vercel added Microsoft's latest speech and transcription models to AI Gateway (Oct 1), so you can call MS TTS and STT through the same endpoint and key as your text models — one less SDK to wire up for voice features.