Apple proposes shared selective persistent memory to retain reusable context in agentic LLM systems

Apple’s Machine Learning Research team proposes shared selective persistent memory, a memory architecture designed to address a core challenge in agentic LLM systems: the loss of reusable context between sessions. In multi-turn tool use scenarios, each session typically starts from zero, discarding configuration choices, domain constraints, data schemas, and tool-use patterns that previously improved productivity. Naively persisting entire conversation histories is both token-inefficient and counterproductive, as irrelevant context can degrade generation quality.

The proposed architecture selectively retains four categories of reusable context—task specifications, data schemas, tool configurations, and output constraints—while discarding session-specific reasoning traces. This selective retention aims to avoid the degradation observed with full-history persistence, which the team notes can actively reduce task completion by biasing the agent with stale reasoning traces.

A key feature of the architecture is its shared nature: workspaces encapsulating selective memory can be transferred across users with role-based access control. This enables collaborative reuse of accumulated context without redundant specification, potentially improving team productivity in enterprise settings.

The team implemented this architecture in a deployed collaborative workspace platform where LLM agents produce, edit, and maintain git-versioned artifacts—including interactive dashboards, structured reports, and data-driven documents—from heterogeneous data sources accessed via multiple connector types such as CSV upload, SQL, REST APIs, and MCP servers. Git-backed versioning with draft isolation allows users to explore modifications risk-free and restore to any prior state without re-invoking the model.

A complementary zero-token data refresh mechanism decouples generated programs from runtime data, enabling artifact reuse without re-invocation. Across three enterprise deployment scenarios, the shared selective persistent memory approach achieved 96% task completion, compared to 79% without memory and 71% with full history persistence. The zero-token refresh mechanism eliminated LLM re-invocation entirely for recurring data updates, resulting in a 14× reduction in task time, while summary-driven generation reduced per-invocation token cost by 97× versus raw data injection.

The team reports that a replication on four public datasets confirmed generalizability, with zero-token refresh succeeding in 12 out of 12 trials. They emphasize that naive full-history persistence degrades task completion, while selective memory outperforms both extremes—demonstrating the value of targeted context retention in agentic systems.