Shared Selective Persistent Memory for Agentic LLM Systems
https://machinelearning.apple.com/research/shared-selective-persistent-memory📌 Apple EMNLP論文：讓AI Agent記住該記的，忘掉該忘的
TL;DR：選擇性共享記憶架構讓agent任務完成率達96%，比完整保留對話記錄還高。
多數人直覺以為，AI agent記得越多對話歷史，表現應該越好。Apple這篇EMNLP論文卻發現：把完整對話歷史都餵給agent，任務完成率反而比完全不給記憶還要低。
🤔 每次對話都從零開始，是agent系統的根本問題
論文指出，透過多輪工具呼叫來生成程式碼的agentic LLM系統，面臨一個結構性問題：每個session都從零開始，先前session中累積的設定選擇、領域限制條件、資料schema與工具使用模式全部被丟棄。但作者也指出，單純把完整對話歷史都持久化保存，既浪費token，也適得其反——不相關的上下文會拉低生成品質。
🧩 只留四種可重複使用的上下文，捨棄過程推理
作者提出「shared selective persistent memory」（選擇性共享持久記憶）架構，只辨識並保留四類可重複使用的上下文：任務規格（task specification）、資料schema、工具設定（tool configuration）、輸出限制條件（output constraint），同時捨棄與特定session相關的推理過程紀錄。這份記憶還是「共享」的：封裝了選擇性記憶的workspace可以透過角色權限控制（role-based access control）在使用者之間傳遞，讓累積下來的上下文能被團隊協作重複使用，不必每次重新描述一遍需求。
這套架構被實作在一個已部署的協作workspace平臺上，LLM agent在其中產生、編輯並維護git版控的成品，包括互動式dashboard、結構化報告與資料驅動文件，資料來源涵蓋CSV上傳、SQL、REST API與MCP伺服器等多種連接器。git-backed版控搭配draft隔離機制，讓使用者可以無風險地嘗試修改，並能還原到任一先前狀態而不必重新呼叫模型。此外，論文提出一個互補的「zero-token資料刷新」機制，把生成的程式與執行期資料解耦，讓成品在資料更新時可以重複使用而不需要重新呼叫LLM。
📊 96% vs 79% vs 71%：記憶不是越多越好
在三個企業部署場景中，選擇性共享持久記憶達到96%的任務完成率，相較於完全不用記憶的79%，以及保留完整對話歷史的71%。zero-token資料刷新機制讓例行性的資料更新完全不需要重新呼叫LLM，任務時間縮短14倍；而摘要驅動（summary-driven）的生成方式，相較於直接注入原始資料，把每次呼叫的token成本降低了97倍。在四個公開資料集上的複製實驗也確認了這套方法的可推廣性，zero-token刷新機制在12次試驗中全數成功。作者特別強調，單純保留完整歷史反而會用過時的推理痕跡誤導agent、實際拉低任務完成率，而選擇性記憶則同時勝過「完全不記」與「全部都記」這兩個極端。
🎯 實務啟示
對於正在建構長期運作的agentic系統的工程師，這篇論文提供了一個明確的設計原則：記憶架構的重點不是「持久化多少」，而是「持久化什麼」。與其把整段對話歷史丟進context，不如明確定義哪幾類資訊（任務規格、schema、工具設定、輸出限制）值得跨session保留，並搭配版控與可還原機制，讓協作與重複使用建立在乾淨的上下文之上，而不是不斷累積的雜訊。
🔗 來源
- 標題：Shared Selective Persistent Memory for Agentic LLM Systems
- 連結：https://machinelearning.apple.com/research/shared-selective-persistent-memory
#AgenticAI #LLM #Apple #EMNLP #AIMemory #MultiTurn #AIAgents #NLProc #MachineLearning #CollaborativeAI
原始資料 Apple ML · 收集於 2026-09-16
摘要原文
research area Speech and Natural Language Processing , research area Tools, Platforms, Frameworks conference EMNLP content type paper published September 2026 Shared Selective Persistent Memory for Agentic LLM Systems Authors Sanjana Pedada, Aditya Dhavala, Neelraj Patil View publication Copy Bibtex Agentic LLM systems that generate code through multi-turn tool use face a fundamental context problem: each session starts from zero, discarding the configuration choices, domain constraints, data schemas, and tool-use patterns that made previous sessions productive. Naively persisting entire conversation histories is both token-inefficient and counterproductive—irrelevant context degrades generation quality. We introduce shared selective persistent memory, a memory architecture for agentic systems that identifies and retains four categories of reusable context—task specifications, data schemas, tool configurations, and output constraints—while discarding session-specific reasoning traces. Crucially, this memory is shared: workspaces encapsulating selective memory can be transferred across users with role-based access control, enabling collaborative reuse of accumulated context without redundant specification. We implement 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 (CSV upload, SQL, REST APIs, and MCP servers). Git-backed versioning with draft isolation enables 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, shared selective persistent memory achieves 96% task completion (vs. 79% without memory and 71% with full history). A complementary zero-token data refresh mechanism eliminates LLM re-invocation entirely for recurring data updates (14×task time reduction), while summary-driven generation reduces per-invocation token cost by 97×versus raw data injection. A replication on four public datasets confirms generalizability, with zero-token refresh succeeding in 12/12 trials. Notably, naive full-history persistence actively degrades task completion by biasing the agent with stale reasoning traces, while selective memory outperforms both extremes. Related readings and updates. Agent Seer: Synthesizing Scenarios from Specification Understanding August 28, 2026 research area Data Science and Annotation , research area Tools, Platforms, Frameworks Evaluating AI agents that use external tools requires realistic test scenarios that capture how practitioners compose tools and iterate across conversation turns. Constructing such scenarios by hand demands deep domain expertise, does not scale across tool ecosystems, and produces static benchmarks that cannot track evolving APIs. We observe that tool specifications—function names, natural-language descriptions, and typed parameter… Read more Less Is More: A Unified Architecture for Device-Directed Speech Detection with Multiple Invocation Types June 8, 2023 research area Methods and Algorithms , research area Speech and Natural Language Processing conference ICASSP Suppressing unintended invocation of the device because of the speech that sounds like wake-word, or accidental button presses, is critical for a good user experience, and is referred to as False-Trigger-Mitigation (FTM). In case of multiple invocation options, the traditional approach to FTM is to use invocation-specific models, or a single model for all invocations. Both approaches are sub-optimal: the memory cost for the former approach grows… Read more Discover opportunities in Machine Learning. Our research in machine learning breaks new ground every day. Work with us
由 claude-code/sonnet 自動生成