arXiv preprint · arXiv:2610.08586 [cs.AI] · First posted · Revised (v3)
MINDSET: Energy-based Memory Evolution for Long Conversational Agents
Preprint · Agentic AI
MINDSET keeps track of changing facts and the conversations that support them.
Authors’ manuscript abstract
A new dialogue turn can replace an earlier fact or introduce one that holds in a different context. Similarity search can return both turns, leaving the answer model to decide which one applies. We introduce MINDSET, a controller that links original dialogue to versioned schemas. A local energy score and stability rules govern reinforcement, supersession, contextual splitting, and schema creation. We evaluate retrieval and answering on a 700-question LoCoMo subset and reuse the controller settings on LongMemEval. In the primary LoCoMo comparison, Recall@8 is 0.6280 for MINDSET and 0.4522 for LightMem on 696 questions with gold evidence (Holm-adjusted p=0.00977). The observed answer-F1 difference over all 700 questions is 0.0256. On all 78 LongMemEval knowledge-update cases, the substitute judge Gemini-3.1-Flash-Lite accepts 54 MINDSET answers and 50 Vector RAG answers. Keeping state history in the index lets retrieval select a short packet of original dialogue for questions about both current and earlier states.
Reproduced from the authors’ manuscript. The arXiv record provides the abstract for the public preprint.
Abstract summary
When someone gives an assistant a second address, they might have moved, or they might be describing where they work. MINDSET uses a local energy score to decide whether a new detail confirms an existing memory, replaces it, applies in a different context, or belongs in a separate memory. Each version points back to the conversation that supplied the evidence. The system can then retrieve what was true at the time a question asks about. In tests on LoCoMo, it ranked relevant evidence higher. The same settings also carried over to questions about changed facts on LongMemEval.
Summarized from arXiv.
Findings
- On 700 LoCoMo questions, MINDSET scored 0.4280 in answer F1 with GPT-OSS-120B.
- Using the same controller settings on LongMemEval-S, the substitute judge Gemini-3.1-Flash-Lite accepted 54 of 78 knowledge-update answers (69.23%).
What the evaluation covers
The memory controller uses hand-tuned settings. The LongMemEval result covers the knowledge-update subset and depends on a substitute model judging the answers. Retrieval also adds latency before the answer model runs.
Sources and citation
Preprint DOI 10.48550/arXiv.2610.08586
View BibTeX citation
@misc{mindset-2026,
title = {{MINDSET: Energy-based Memory Evolution for Long Conversational Agents}},
author = {Sujato Dutta and Sreekruthy Tummala and Shashank Vanga and Ayushmi Pavani and BHVSP Subrahmanyam and Neeraj Choudhary},
year = {2026},
eprint = {2610.08586},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
doi = {10.48550/arXiv.2610.08586},
url = {https://arxiv.org/abs/2610.08586}
}