C3M: Cross-Session Multimodal Memory Maintenance for Long-Horizon Tasks

Source facts

The preprint proposes bounded query-blind memory for long-horizon tasks. It describes relation-aware updates for complementary and incompatible records, budgeted routing that selects index pages and expands source evidence under a reader budget, and compact multimodal memory that preserves temporal distinctions and source links. The abstract indicates that code is available through the project link.

Analysis

The concept treats agent memory as an update-and-provenance problem rather than a single growing context window. Preserving source links and temporal distinctions could make long-running systems more auditable, if the routing and maintenance policy remain reliable under changing tasks.

Uncertainty/Doubt

• This is an arXiv preprint. The memory-quality results, code, and behavior under adversarial or highly contradictory histories were not independently audited here.