Research
Shared Selective Persistent Memory for Agentic LLM Systems
Apple paper: keep specs, schemas, tool config, and constraints; drop session traces. Selective shared memory beat both no memory and full history.
Sanjana Pedada, Aditya Dhavala, and Neelraj Patil (Apple Machine Learning Research, September 2026) argue that agent sessions should not start from zero, and should not keep the whole transcript either.
Claim
Retain four reusable categories — task specifications, data schemas, tool configurations, and output constraints — and discard session-specific reasoning traces. Workspaces of that memory can be shared across users with role-based access. Artifacts are git-versioned with draft isolation. A zero-token data refresh updates programs without calling the model again.
Reported results
- 96% task completion with selective memory, versus 79% with no memory and 71% with full history
- Full-history persistence degraded completion by biasing the agent with stale traces
- Zero-token refresh removed re-invocation on recurring data updates (14× task-time reduction in their enterprise scenarios); replication succeeded on 12/12 public-dataset trials
- Summary-driven generation cut per-invocation tokens 97× versus injecting raw data
"Selective memory outperforms both extremes."