Apple Unveils Shared Selective Persistent Memory Architecture for Agentic LLMs

Apple researchers have unveiled a shared selective persistent memory architecture to address the fundamental context issues in agentic LLM systems that generate code through multi-turn tool use. Existing systems had to relearn settings, domain constraints, data schemas, etc., from scratch every session, while storing the entire conversation history led to token inefficiency and degraded generation quality.

The new architecture identifies and maintains four reusable categories of context (task specifications, data schemas, tool configurations, output constraints) and discards per-session reasoning traces. This memory can be shared among users via role-based access control (RBAC), enabling collaborative reuse of accumulated context.

In the implemented collaborative workspace platform, LLM agents generate and maintain Git version-controlled artifacts (dashboards, reports, etc.) through various sources such as CSV, SQL, REST APIs, and MCP servers. The zero-token data refresh mechanism separates generated programs from runtime data, completely eliminating LLM re-invocations during repetitive data updates.

In three enterprise deployment scenarios, this architecture achieved a 96% task completion rate. This outperforms the 79% achieved without memory and the 71% achieved when storing full history. The zero-token refresh mechanism realized a 14x reduction in task time during repetitive data updates, and summary-based generation achieved a 97x reduction in token costs compared to raw data injection. Replication experiments conducted on four public datasets also demonstrated generalizability, with zero-token refresh achieving 12/12 success.