From Memory Hits to Answerable Evidence: Query-Conditioned Memory Compilation for Long-Term Agent Memory

TL;DR

MemCompile introduces a query-driven method for organizing long-term conversational memory by first identifying relevant memory fragments through compact descriptors and then compiling them into a focused, answer-ready context. It separates memory addressing from evidence presentation, reducing redundancy and improving context quality by dynamically allocating a fixed evidence budget across stable memory units.

Why It Matters

This approach enhances the reliability and efficiency of AI agents in sustained conversations by ensuring that only relevant, well-structured evidence is used for responses, reducing hallucination risks and improving answer accuracy over time.

Key Findings

• MemCompile improves Judge Accuracy to 80.52% on the LoCoMo benchmark, outperforming prior methods.
• The framework achieves better quality-cost trade-offs by limiting redundant or incomplete context through bounded evidence allocation.
• Performance gains generalize across benchmarks, including LongMemEval-S, indicating broad applicability to long-term memory systems.