Copy the memory, not the permission: a test case for AI agent obligations
tallmetommy · x · 2026-11-09
The author proposes a concrete thought experiment exposing a core tension in agent design: memory can be copied, permissions cannot. Give a fictional worker a $400 debt, queue a payroll deduction, take payroll offline, cancel the debt on appeal, restart payroll — she holds a letter saying she won; does her pay stub agree?
A sharper analogy follows: copy an AI that knows your secrets and has permission to make one payment. Both copies must keep the secrets, but they cannot both make the payment. The takeaway: copying memory must preserve obligations without multiplying permissions — a usable criterion for agent memory sync, approval flows, and eval test scenarios.
More from AGI Musings
- Perplexity CEO touts model-neutrality as crypto's 'credible neutrality' idea reaches AI — dnlkwk · 2026-11-09
- InstaDeep CEO: Anthropic's unreleased Models 3 and 4 sit halfway to AI2027's Agent-2 — connoraxiotes · 2026-11-09
- Nature review: AI has transformed maths — and may outgrow human insight — LucaAmb · 2026-11-09
- 93% of Chinese think AI will improve their lives — vs 36% of Americans, 0% of Europeans — alexmacgregor__ · 2026-11-09
- $2.5B ARR per capability point: Ben Todd rebuts the AI slowdown argument — ronbodkin · 2026-11-09
- Writing is a thinking tool: without paper you're just a finite automaton — DynamicWebPaige · 2026-11-09