Source: arXiv (Apple Inc.) — 2026-07-10
Summary
Apple researchers propose a memory architecture for agentic LLM systems that selectively persists four narrow categories of reusable context — task specs, data schemas, tool configurations, and output constraints — across sessions, while deliberately discarding the model's raw reasoning traces. Across three enterprise scenarios, the approach reports 96% task completion versus 79% with no memory and 71% with full conversation-history replay, plus a roughly 14x reduction in task time via what the paper calls "zero-token refresh."
Key Takeaways
- The core bet is that most of what makes agent memory expensive and noisy is the reasoning trace, not the facts — so the system keeps the facts and throws away the thinking.
- Full-history replay (71% completion) actually underperformed no memory at all (79%) in their tests, suggesting naive "just remember everything" memory can hurt more than it helps by re-injecting stale or irrelevant reasoning.
- Selective persistence hit 96% completion, a meaningfully wide margin over both baselines, while cutting task time by roughly 14x.
- "Zero-token refresh" — reloading only structured facts instead of replaying prior conversation — is the mechanism behind the speedup, since the agent isn't re-reading and re-reasoning over its own past thought process.
Reel Script
Hook (17s / 38 words) Most AI agent memory systems make agents slower and dumber the more they remember. Apple just published a fix that's almost embarrassingly simple: stop saving the thinking, only save the facts.
Core Concept (70s / 155 words) Here's the problem this solves. When an AI agent works across multiple sessions, the obvious way to give it memory is to save its past conversation and replay it next time — like handing someone their own diary before they start a new task. Apple's paper shows that actually makes things worse: agents given full conversation history scored 71% task completion, worse than agents given no memory at all, which hit 79%. Why? Because the diary is full of dead-end reasoning, wrong guesses, and stuff that was only relevant in the moment. Their fix keeps four specific things instead: what the task actually specifies, what the data looks like, how the tools are configured, and what the output has to satisfy. Everything else — the "thinking out loud" — gets thrown away. It's the difference between handing someone a clean reference sheet versus handing them your entire messy scratch pad from last time.
Hands-On (65s / 145 words) The numbers are the artifact here. No memory: 79% task completion. Full conversation-history replay: 71% — worse, not better. Selective persistent memory, keeping only those four structured categories: 96% completion. That's a 17-point jump over the no-memory baseline and a 25-point jump over naive full-history replay. On top of accuracy, they report roughly a 14x reduction in task time, which comes from what they call zero-token refresh — instead of re-feeding the model a long transcript it then has to re-read and re-reason over, you just load the structured facts directly, so there's far less to process before the agent can act. If you're building any agent that needs to remember something across sessions, this is a strong argument for scoping exactly what "memory" means before you build it.
Takeaway (25s / 55 words) If your agent's memory system is "just save the transcript and replay it," these numbers say you're probably making it slower and less reliable, not more capable. Go look at what your agent actually needs to remember — task specs and schemas, not its own internal monologue.