Shared Selective Persistent Memory for Agentic LLM Systems
Naively persisting entire conversation histories is token-inefficient and counterproductive: irrelevant context degrades generation quality. We introduce shared selective persistent memory, an architecture 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 without redundant specification.
- ▪Naively persisting entire conversation histories is token-inefficient and counterproductive: irrelevant context degrades generation quality.
- ▪We introduce shared selective persistent memory, an architecture that identifies and retains four categories of reusable context (task specifications, data schemas, tool configurations, and output constraints) while discarding session-speci
- ▪Crucially, this memory is shared: workspaces encapsulating selective memory can be transferred across users with role-based access control, enabling collaborative reuse without redundant specification.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2607.09493 (cs) [Submitted on 10 Jul 2026] Title:Shared Selective Persistent Memory for Agentic LLM Systems Authors:Sanjana Pedada, Aditya Dhavala, Neelraj Patil View a PDF of the paper titled Shared Selective Persistent Memory for Agentic LLM Systems, by Sanjana Pedada and 2 other authors View PDF HTML (experimental) Abstract: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 token-inefficient and counterproductive: irrelevant context degrades generation quality.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
