STITCH: Agentic Memory System for LLMs in Long-Horizon Tasks
Researchers have introduced STITCH (Structured Intent Tracking in Contextual History), an innovative memory system aimed at enhancing the performance of large language models in lengthy, goal-driven interactions. This system tackles the challenge of retrieving contextually mismatched information when similar entities and facts appear under varying latent goals. STITCH organizes each trajectory step with a structured retrieval cue and contextual intent, retrieving history by aligning with the intent of the current step. Contextual intent encompasses the present latent goal, action type, and significant entity types. During the inference phase, STITCH prioritizes and filters memory snippets based on intent compatibility, minimizing semantically similar yet contextually incompatible history. Additionally, the team has developed CAME-Bench for assessment. The paper can be found on arXiv with ID 2601.10702.
Key facts
- STITCH stands for Structured Intent Tracking in Contextual History.
- It is an agentic memory system for large language models.
- It indexes each trajectory step with a structured retrieval cue and contextual intent.
- Contextual intent includes latent goal, action type, and salient entity types.
- STITCH filters memory snippets by intent compatibility during inference.
- CAME-Bench is introduced for evaluation.
- The paper is on arXiv with ID 2601.10702.
- The system aims to reduce interference from context-mismatched history.
Entities
Institutions
- arXiv