Apple's shared selective persistent memory hits 96% task completion

The architecture is built around a problem Apple's researchers describe plainly: in multi-turn tool use, each session starts from zero, discarding the configuration choices, domain constraints, data schemas, and tool-use patterns that made previous sessions productive. Their answer is to keep only part of the past, identifying four categories of reusable context while discarding session-specific reasoning traces.

The platform side of the work is a deployed collaborative workspace where LLM agents produce, edit, and maintain git-versioned artifacts — interactive dashboards, structured reports, and data-driven documents — from heterogeneous sources accessed via CSV upload, SQL, REST APIs, and MCP servers. Git-backed versioning with draft isolation lets users explore modifications risk-free and restore to any prior state without re-invoking the model, and a zero-token data refresh decouples generated programs from runtime data so artifacts can be reused without re-invocation.

The standout comparison is against the obvious baseline. Naive full-history persistence actively degrades task completion, the team reports, apparently biasing the agent with stale reasoning traces, while selective memory outperforms both extremes. The reading is that the commercial benefit hinges on what gets filtered out, not on how much context is retained — though the 96% figure comes from three enterprise deployment scenarios, so whether it generalizes rests on the replication across four public datasets.