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

TL;DR - C3M is a bounded multimodal memory system for long-horizon, cross-session tasks that preserves source text-image evidence and provenance. It aims to prevent compression from erasing visual details or merging observations that appear similar but are incompatible.

• Maintains a compact active index linked to persistent source evidence.
• Uses relation-aware updates to merge safe redundancies while retaining complementary or conflicting records.
• Routes queries to relevant index pages, then expands linked evidence within a fixed reader budget.
• Preserves temporal distinctions and source links needed for reliable downstream reasoning.