Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2607.09493.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

Researchers propose shared selective persistent memory for agentic LLM systems, retaining four categories of reusable context (task specifications, data schemas, tool configurations, output constraints) while discarding session-specific reasoning. Implemented in a collaborative workspace platform, it achieves 96% task completion versus 79% without memory and 71% with full history.

Findings contradict naive RAG assumptions: full history persistence degrades agent performance. Selective four-category memory plus zero-token refresh cuts re-invocation 14x and token cost 97x in tested scenarios.

Cited by

No citations on record.