E-mem: Multi-agent Episodic Memory for LLM Reasoning
Researchers have introduced E-mem, a novel framework for Large Language Model (LLM) agents that transitions from traditional memory preprocessing to the reconstruction of episodic contexts. Drawing inspiration from biological engrams, E-mem features a diverse hierarchical system where several assistant agents retain uncompressed memory contexts, while a central master agent coordinates overall planning. This method seeks to maintain the integrity of context for System 2 reasoning, circumventing the harmful de-contextualization that results from compressing intricate sequential dependencies into predetermined formats like embeddings or graphs. The framework allows assistants to reason locally within activated segments and extract relevant context. The paper can be found on arXiv with ID 2601.21714.
Key facts
- E-mem is a framework for LLM agent memory.
- It shifts from Memory Preprocessing to Episodic Context Reconstruction.
- Inspired by biological engrams.
- Uses heterogeneous hierarchical architecture.
- Multiple assistant agents maintain uncompressed memory contexts.
- Central master agent orchestrates global planning.
- Assistants locally reason within activated segments.
- Paper ID: arXiv:2601.21714.
Entities
Institutions
- arXiv