RaMem: Contextual Reinstatement for Long-term Agentic Memory

Modern LLM agents often store long-term memories as compressed, reusable fragments. While this makes information easier to store and retrieve, it leads to a phenomenon the authors call context collapse. In this state, an agent might retrieve a memory that is semantically relevant to a query but factually invalid because it originated from a different session, time, or situational context. Current systems struggle to distinguish between these related-but-invalid memories and the true evidence required for the task.

To solve this, the authors introduce Contextual Reinstatement for Agentic Memory (RaMem). Instead of treating memories as isolated text, RaMem treats them as situated evidence. The framework operates in four stages:

1. Evidence Anchoring: Each memory is tagged with metadata, including event time, session span, and participants.
2. Recall Condition Induction: The system analyzes the user query to identify the specific context required (e.g., a specific date or person).
3. Validity-Aware Retrieval: The system prioritizes memories that match the query's contextual requirements, using content-based retrieval only as a fallback.
4. Context-Preserved Synthesis: The final evidence is presented to the LLM with its structured context intact, ensuring the generator can reason about the validity of the information.

Experiments on long-term memory benchmarks (LoCoMo) demonstrate that RaMem consistently outperforms strong baselines, achieving average F1 score gains of over 10% across multiple model backbones. This suggests that for long-term agentic tasks, the ability to verify the context of a memory is just as critical as the ability to retrieve its content. By moving beyond simple relevance-based retrieval, RaMem enables more reliable reasoning in complex, multi-session environments.