---
title: "Shared Selective Persistent Memory for Agentic LLM Systems"
description: "Agentic LLM systems that generate code through multi-turn tool use lose configuration choices, domain constraints, schemas, and tool-use patterns when each session starts from zero. The proposed shared selective persistent memory retains task specifications, data schemas, tool configurations, and output constraints while discarding session-specific reasoning traces, and allows workspaces to be transferred across users with role-based access control. In a deployed collaborative workspace platform, agents create and maintain git-versioned dashboards, reports, and data-driven documents from CSV, SQL, REST API, and MCP sources; draft isolation supports risk-free changes and restoration without re-invoking the model. Results across three enterprise scenarios show 96% task completion, compared with 79% without memory and 71% with full history, while zero-token refresh cut recurring-update task time 14× and summary-driven generation reduced token cost 97× versus raw data injection."
---

# Shared Selective Persistent Memory for Agentic LLM Systems

[Apple](https://yomu.fyi/company/apple) · Sep 16, 2026

**Type:** Benchmark

## Summary

Agentic LLM systems that generate code through multi-turn tool use lose configuration choices, domain constraints, schemas, and tool-use patterns when each session starts from zero. The proposed shared selective persistent memory retains task specifications, data schemas, tool configurations, and output constraints while discarding session-specific reasoning traces, and allows workspaces to be transferred across users with role-based access control. In a deployed collaborative workspace platform, agents create and maintain git-versioned dashboards, reports, and data-driven documents from CSV, SQL, REST API, and MCP sources; draft isolation supports risk-free changes and restoration without re-invoking the model. Results across three enterprise scenarios show 96% task completion, compared with 79% without memory and 71% with full history, while zero-token refresh cut recurring-update task time 14× and summary-driven generation reduced token cost 97× versus raw data injection.

## Context

The system addresses repeated specification and context loss in multi-turn, tool-using coding sessions. It is deployed in a collaborative workspace platform where agents work with heterogeneous data sources and produce git-versioned artifacts.

## Approach / What changed

The architecture selectively persists task specifications, data schemas, tool configurations, and output constraints, while excluding session-specific reasoning traces. Shared workspaces use role-based access control, git-backed versioning, and draft isolation. A zero-token refresh mechanism separates generated programs from runtime data, and summary-driven generation avoids injecting raw data into model prompts.

## Takeaways

- Selective memory achieved 96% task completion across three enterprise deployment scenarios, versus 79% without memory and 71% with full-history persistence.
- Full-history persistence can reduce performance because stale reasoning traces bias the agent; selective memory outperformed both no-memory and full-history approaches.
- Zero-token data refresh succeeded in 12/12 public-dataset trials, while summary-driven generation reduced per-invocation token cost by 97× versus raw data injection.

## FAQ

### What does the shared memory retain, and what does it discard?

It retains task specifications, data schemas, tool configurations, and output constraints. Session-specific reasoning traces are discarded to avoid carrying irrelevant or stale context between sessions.

### How can multiple users reuse the accumulated context?

Selective memory is packaged in workspaces that can be transferred across users. Role-based access control governs that shared workspace access.

### How do recurring data updates avoid another model invocation?

The zero-token data refresh mechanism decouples generated programs from runtime data, allowing artifacts to be reused without re-invoking the model. It produced a 14× task-time reduction and succeeded in all 12 public-dataset trials.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Architecture](https://yomu.fyi/topic/architecture), [LLMs](https://yomu.fyi/topic/llm)

- Source: [Apple](https://machinelearning.apple.com/research/shared-selective-persistent-memory)
- Source URL: https://machinelearning.apple.com/research/shared-selective-persistent-memory
- Ingested by Yomu: 2026-09-16T20:00:52.302Z

[Read original post](https://machinelearning.apple.com/research/shared-selective-persistent-memory)