AuthorsJohannes Kirmayr, Lukas Stappen, Phillip Schneider et al.
TL;DR
CarMem uses category-bound preference extraction plus LLM-driven maintenance to achieve F1 between 0.78 and 0.95 on CARMEM while cutting redundant preferences by 95%.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CarMem targets the issue that many systems struggle with retaining user preferences, leading to repetitive user requests and disengagement.
In in-car assistants, unbounded preference extraction conflicts with GDPR data minimization and EU AI Act transparency, causing privacy, trust, and resource problems.
HOW IT WORKS
CarMem centers on Preference Extraction, Preference Maintenance, and Preference Retrieval, all constrained by a hierarchical, predefined category schema for in-car preferences.
You can think of CarMem like a card catalog for car-related preferences, where each card is filed under specific shelves such as Vehicle Settings and Comfort or Entertainment.
By bounding extraction to actionable categories, CarMem enables transparent, opt-out friendly long-term memory that a plain context window cannot provide without violating data minimization.
DIAGRAM
This diagram shows how CarMem processes multi turn in car conversations to extract, maintain, and later retrieve preferences.
DIAGRAM
This diagram shows how CARMEM is generated and then used to benchmark extraction, maintenance, and retrieval in CarMem.
PROCESS
Preference Extraction
CarMem runs Preference Extraction with LLM function calling over the conversation, filling a hierarchical schema of main, sub, and detail categories.
Preference Maintenance
CarMem applies Preference Maintenance functions pass, update, and append, comparing incoming preferences with existing ones per detail category.
Preference Retrieval
CarMem performs Preference Retrieval by embedding detail category, attribute, and sentence, then retrieving top n preferences per subcategory for a new utterance.
Personalized Response Generation
CarMem supplies retrieved preferences to the in car assistant, enabling personalized responses aligned with vehicle functions and user opt out choices.
KEY CONTRIBUTIONS
Category-bound preference memory system
CarMem introduces a category-bound preference memory system with Preference Extraction, Preference Maintenance, and Preference Retrieval tailored to in car assistants, reaching detail-level F1 of 0.78.
Closed-world in-car conversational dataset CARMEM
CarMem provides the synthetic closed world dataset CARMEM with 1,000 Extraction Conversations, 1,000 Retrieval Utterances, and 3,000 Maintenance Utterances for systematic memory evaluation.
LLM-driven maintenance with strong redundancy reduction
CarMem extends LLM driven maintenance with function calling, achieving a 95% reduction in redundant preferences and 93% reduction in contradictory preferences in Maintenance Utterances.
RESULTS
Detail F1
0.78 F1
detail level precision 0.75 and recall 0.81 in 41 categories
Main F1
0.94 F1
main level precision 0.93 and recall 0.94 across 4 categories
Redundant reduction
95% reduction
reduction in redundant preferences via maintenance compared to naive append
Optimal retrieval
0.87 accuracy
top n accuracy using enriched embeddings with detail category and attribute
On the CARMEM dataset, CarMem is evaluated for extraction, maintenance, and retrieval in an in car voice assistant setting. These results show that CarMem can reliably structure preferences, keep storage consistent, and retrieve relevant memories under a realistic closed world scenario.
BENCHMARK
On the CARMEM dataset, CarMem is evaluated for extraction, maintenance, and retrieval in an in car voice assistant setting. These results show that CarMem can reliably structure preferences, keep storage consistent, and retrieve relevant memories under a realistic closed world scenario.
BENCHMARK
Micro averaged F1 scores for preference extraction at main, sub, and detail category levels.
KEY INSIGHT
CarMem’s Out of Schema experiment shows that when a subcategory is excluded, non extraction is achieved 75% of the time, despite complex conversations.
This is surprising because LLMs often over extract, yet CarMem’s function schema strongly bounds extraction, contradicting the assumption that LLM memories must be unstructured and noisy.
WHY IT MATTERS
CarMem unlocks controllable, category bound long term memory for voice assistants, aligning personalization with GDPR and EU AI Act transparency requirements.
Builders can now deploy in car assistants that remember actionable preferences, let users opt out per category, and maintain clean memory stores without bespoke infrastructure.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…