Beyond Marginal Influence: Interaction-Aware Memory Selection for LLM Agents
JointCore finds jointly sufficient LLM-agent memories, recovering minimum-cost sets in all 6,912 exact benchmark cases.
The paper studies per-decision memory sufficiency for LLM agents, seeking small subsets that preserve full-memory behavior when evidence is redundant or jointly necessary. JointCoreBench supplies exact sufficient-set oracles, and cost-ordered JointCore recovers a minimum-cost set in 100% of 6,912 exact conditions while cutting evaluations from 30.00 to 11.13. Across four 12B–70B model families, it averages 10.28 queries versus 16.00 for singleton deletion; the authors present it as a benchmark baseline, not production memory compression.
- JointCore selects jointly sufficient memory subsets rather than ranking records independently.
- JointCoreBench covers nine interaction families and six workflow domains.
- Cost-ordered JointCore recovered a minimum-cost set in all 6,912 exact conditions.
- Subset evaluations fell from 30.00 to 11.13, a 62.9% reduction.
- Authors say it is a benchmark baseline, not deployment-ready compression.
From the original article
1 min read at arxiv.orgLLM-agent memory selection often uses retrieval rankings based on record relevance or individual influence. Such rankings cannot identify sufficient subsets under redundant or jointly necessary evidence. We study per-decision memory sufficiency: selecting a small subset that preserves full-memory behavior.
By BHVSP Subrahmanyam · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.