The Pulse
Apple Reports 96% Completion With Selective Memory, Versus 71% With Full History
Apple Machine Learning Research reports that selective persistent memory outperformed both no memory and full conversation histories in three enterprise deployment scenarios, while a zero-token refresh mechanism completed all 12 trials on f
AI.info Team ·
Apple Machine Learning Research reports that an agent using selective persistent memory completed 96% of tasks across three enterprise deployment scenarios. The same system completed 79% of tasks without memory and 71% when given the full conversation history.
The finding comes from Shared Selective Persistent Memory for Agentic LLM Systems, a paper by Sanjana Pedada, Aditya Dhavala and Neelraj Patil. The Apple research page says the work was published in September 2026 and describes a memory architecture for agentic systems that generate code through multi-turn tool use.
The paper’s authors write:
Naively persisting entire conversation histories is both token-inefficient and counterproductive—irrelevant context degrades generation quality.
The proposed architecture retains four categories of reusable context: task specifications, data schemas, tool configurations and output constraints. It discards session-specific reasoning traces instead of carrying the entire prior conversation into a new session.
Task specifications preserve requirements that can be reused across sessions. Data schemas capture information about the structure of source data. Tool configurations describe the connectors and tools available to the agent, while output constraints preserve requirements governing the artifacts that the system produces.
The memory is shared across users through workspaces with role-based access control. The system is implemented in a collaborative workspace platform where agents produce, edit and maintain git-versioned artifacts, including interactive dashboards, structured reports and other data-driven documents.
The platform supports heterogeneous data sources accessed through CSV uploads, SQL, REST APIs and MCP servers. Git-backed versioning provides draft isolation, allowing users to test changes and restore an earlier state without invoking the model again.
Refreshes without another model invocation
The paper also describes a zero-token data-refresh mechanism. Generated programs remain separate from runtime data, allowing an existing artifact to use compatible updated data without another LLM invocation.
Apple reports a 14× reduction in task time from this refresh mechanism. A replication using four public datasets succeeded in all 12 trials, according to the research page.
For tasks that do require model input, the system uses summaries rather than injecting raw data. Apple reports that summary-driven generation reduces per-invocation token cost by 97× compared with raw-data injection.
Why full histories performed worse
The research page says that full-history persistence can actively reduce task completion by biasing the agent with stale reasoning traces. Earlier reasoning, rather than only reusable requirements and data structure, may introduce context that is no longer appropriate to the current task.
The results do not establish that conversation history should always be discarded. They support a narrower design choice: preserve information that can be reused, while leaving behind session-specific reasoning and other context that may misdirect a later generation.
Apple’s reported comparison therefore favors selective memory over both extremes tested on the page. Agents with no persistent memory achieved lower completion than the selective-memory system, while agents given complete histories performed worse still. The accompanying refresh design also reduced the need to call the model for recurring data updates.