Apple researchers publish EMNLP paper on 'shared selective persistent memory' architecture for agentic LLM systems, reporting 96% task completion in enterprise tests

Apple's ML research team published an EMNLP paper on "shared selective persistent memory," an architecture that keeps task specs, data schemas, tool configs and output constraints between agent sessions while discarding raw reasoning traces, with role-based sharing across users. Testing in an internal collaborative workspace tool plus four public datasets reported 96% task completion versus 79% with no memory and 71% with naive full-history retention, a 14x speedup on repeat data-refresh tasks, and 97x lower per-call token cost than dumping raw data into the prompt each time.

The most useful finding here is negative: keeping everything (full session history) performed worse than keeping nothing, which is a real warning for anyone building agent memory today, since "just persist more context" is a common default design choice. Beyond that, nothing has shipped. No code, weights, or product release accompanies the paper, so the only people who can act on this are researchers and engineering teams willing to reimplement the described architecture from the paper's description alone.

The headline numbers (96%, 14x, 97x) come entirely from Apple's own internal enterprise deployments, which are unverifiable by outsiders, and from a self-run replication on public datasets rather than independent testing. Selective agent memory is already a crowded research area, so treat this as one data point among many similar efforts rather than a breakthrough, despite the clean-looking stats.

1. 01Whether Apple or others release code, a reference implementation, or benchmark data that lets outsiders reproduce the 96%/79%/71% comparison
2. 02Whether this architecture (or a close variant) shows up in an actual Apple product, API, or developer tool within the next 6-12 months
3. 03Whether independent teams building agentic systems cite or replicate the full-history-degrades-performance finding, which is the paper's most durable and testable claim