HiCoMER
A hierarchical memory retrieval framework designed for multi-agent collaborative scenarios. It addresses the problem of semantically relevant but outdated memories being incorrectly retrieved by first assessing memory validity before executing retrieval. The framework comprises three components: a hierarchical memory conflict updater, a validity-aware retriever, and a memory-anchored answer generator.
Timeline (last 90 days)
HiCoMER is a memory-augmented framework for multi-agent collaborative scenarios, whose key design is to first maintain the validity of team memory and individual memory, and then retrieve from memories that are still valid
The HiCoMER framework consists of three modules: a hierarchical memory conflict updater, a validity-aware retriever, and a memory-anchored answer generator
The hierarchical memory conflict updater is responsible for maintaining the validity of team and individual memories, and handles conflicts between individual memories and team consensus
The validity-aware retriever retrieves only from memories that are still valid, rather than performing full retrieval
The hierarchical memory conflict updater manages memory by attaching validity flags to memory entries, marking individual memory entries that contradict new consensus as invalid rather than deleting them outright when the team reaches a new consensus
HiCoMER decouples the two questions of 'what memories still hold' and 'which memory is most relevant,' solving them independently in separate stages
HiCoMER regards team memory as representing the current collective consensus and thus having higher authority, while individual memory may contain intermediate states that have not yet converged or have already been overturned
The research team constructed two new datasets oriented toward collaborative scenarios and based on memory-grounded QA to evaluate HiCoMER
In experiments on the two datasets, HiCoMER maintains a consistent advantage over multiple strong baseline methods, specifically reflected in three aspects: significantly reducing retrieval of outdated memories, better preserving team consensus, and improving QA quality
The two datasets of HiCoMER are designed to cover agent teams of different scales and different frequencies of memory updates
1 more timeline events
All Facts (11)
HiCoMER is a memory-augmented framework for multi-agent collaborative scenarios, whose key design is to first maintain the validity of team memory and individual memory, and then retrieve from memories that are still valid
50%UnverifiedThe HiCoMER framework consists of three modules: a hierarchical memory conflict updater, a validity-aware retriever, and a memory-anchored answer generator
50%UnverifiedThe hierarchical memory conflict updater is responsible for maintaining the validity of team and individual memories, and handles conflicts between individual memories and team consensus
50%UnverifiedThe validity-aware retriever retrieves only from memories that are still valid, rather than performing full retrieval
50%UnverifiedThe hierarchical memory conflict updater manages memory by attaching validity flags to memory entries, marking individual memory entries that contradict new consensus as invalid rather than deleting them outright when the team reaches a new consensus
50%UnverifiedHiCoMER decouples the two questions of 'what memories still hold' and 'which memory is most relevant,' solving them independently in separate stages
50%UnverifiedHiCoMER regards team memory as representing the current collective consensus and thus having higher authority, while individual memory may contain intermediate states that have not yet converged or have already been overturned
50%UnverifiedThe research team constructed two new datasets oriented toward collaborative scenarios and based on memory-grounded QA to evaluate HiCoMER
50%UnverifiedIn experiments on the two datasets, HiCoMER maintains a consistent advantage over multiple strong baseline methods, specifically reflected in three aspects: significantly reducing retrieval of outdated memories, better preserving team consensus, and improving QA quality
50%UnverifiedThe two datasets of HiCoMER are designed to cover agent teams of different scales and different frequencies of memory updates
50%UnverifiedHiCoMER was proposed as an arXiv paper
50%