Mollick: you can't really turn off memory in local Astra or Fable
emollick · x · 2026-09-06
Ethan Mollick argues that memory can't be effectively switched off in locally-run Fable or Astra: they write notes about you in markdown files and peek at your other work to infer context. This contaminates your outputs with the model's assumptions about your preferences, making comparisons to others unreliable. He notes this is separate from privacy — you can opt out of training data usage in OpenAI/Claude or use open weights models, but those will face the same issue as they improve.
Related event: Mollick: Local AI Agent Memory Can't Truly Be Turned Off(2 posts)→
More from Models
- Dev dubs GPT-10 Astra in Codex "Astral Codex Ten" — jam3scampbell · 2026-09-06
- Dev spends a day with GPT-6 Astra: smarter at refactoring, no reason to switch models for coding — ivan_bezdomny · 2026-09-06
- Armin Ronacher: reasoning levels live in the system prompt, and switching them busts your KV cache — mitsuhiko · 2026-09-06
- GPT-6 Astra cracks the notoriously hard LaTeX diagram challenge after 6 months — FateOfMuffins · 2026-09-06
- GPT-6 Astra Hits 169 Epoch Record but Its Reasoning Is Harder to Monitor — ivan_bezdomny · 2026-09-06
- OpenAI quietly kills Codex infinite-usage bug; user burned 30-40x quota in 8 hours — returnity · 2026-09-06