CABLE: Extending the Reach of Memory Retrieval via Complementary Antecedent-Based Linking and Expansion

Agent memory systems heavily rely on semantic similarity to organize and retrieve memories. This works well for topical recall, but it often misses semantically distant memories that implicitly underlie later events, such as earlier causes, plans, or motivations. Existing memory graphs partially address cross-memory structure, yet many of their links remain driven by semantic overlap and therefore duplicate what the base retriever can already recover. We argue that explicit links should instead prioritize retriever-complementary associations. We present CARMA (\textbf{C}omplementary \textbf{A}ntecedent \textbf{R}easoning for \textbf{M}emory \textbf{A}ssociation), a plug-in augmentation that adds links outside the base retriever's semantic reach. For each new memory, CARMA generates antecedent-oriented queries, retrieves prior memories, subtracts those already reachable by direct semantic search, and verifies the remaining candidates before inserting them into a sparse directed graph. At inference time, CARMA expands the host system's retrieved seeds along these links to surface implicit supporting evidence. We integrate CARMA into A-MEM and SimpleMem and evaluate on LoCoMo and MA-LongMemEval with Qwen3.5-27B, DeepSeek-chat, and GPT-4o-mini. CARMA improves LLM-judge accuracy in every evaluated setting, with the largest gains on question types that require implicit cross-memory association, including open-domain, multi-session, and preference-oriented questions. These results suggest that memory graphs are most useful when they prioritize semantically unrecoverable yet explanatorily relevant associations.