Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.02613.

Pipeline notes

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

MemArena benchmark for on-device personal memory assistants simulates 50 agents over 15 days, evaluating five open-weight LLMs with BM25-RAG, Oracle, Memobase, and MemSearch backends across six recall, reasoning, and trustworthiness dimensions. Key findings: backend choice outweighs reader scaling, permission-aware access fails universally, and search adds modest 7–87 ms latency on edge hardware.

Concrete benchmark numbers on on-device memory backends with an accompanying simulator release, directly useful for builders of private, edge-deployed agentic assistants.

Cited by

No citations on record.