# LoCoMo-10

The work proposes Gated Memory, a formation layer interposed between conversation and storage: an utterance-level admission gate evaluates candidate facts before extraction, and admitted facts undergo entity scope classification and conditional enrichment, with only the current exchange treated as a fact source and prior turns used read-only for reference resolution. On LoCoMo-10, changing only what enters the memory store while holding retrieval, generation, and evaluation fixed yields a +2.6% overall relative gain in LLM-judge accuracy, with Single-Hop +4.4%, Open Domain +3.3%, Temporal +1.7%, and a 4.2% Multi-Hop regression.

The work proposes Gated Memory, a formation layer interposed between conversation and storage: an utterance-level admission gate evaluates candidate facts before extraction, and admitted facts undergo entity scope classification and conditional enrichment, with only the current exchange treated as a fact source and prior turns used read-only for reference resolution. On LoCoMo-10, changing only what enters the memory store while holding retrieval, generation, and evaluation fixed yields a +2.6% overall relative gain in LLM-judge accuracy, with Single-Hop +4.4%, Open Domain +3.3%, Temporal +1.7%, and a 4.2% Multi-Hop regression.