Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents

Researchers propose DCPM, a dual-process cognitive memory system for LLM agents that organizes memory hierarchically from raw inputs to cross-domain patterns. The system uses a synchronous writer to record belief revisions and an asynchronous engine to induce schemas and detect cross-domain patterns, achieving significant improvements on personalization benchmarks requiring implicit reasoning about user evolution.

This research addresses a fundamental limitation in current LLM agent architectures: the inability to maintain coherent, evolving long-term memory that supports implicit personalization. Traditional memory systems flatten complex cognitive processes into single retrieval surfaces optimized for surface-level recall, failing to capture how users and contexts evolve over time. DCPM introduces architectural innovation inspired by dual-process cognitive theory, implementing two distinct mechanisms that operate at different timescales and abstraction levels.

The System1 component functions as a synchronous daytime writer, recording belief revisions through doubly-linked supersedes chains that maintain a diachronic record of how facts and beliefs change. System2 operates asynchronously as a nighttime consolidation engine, performing schema induction and cross-domain pattern detection to build higher-level abstractions. This separation mirrors how human memory consolidation works, where immediate recording differs fundamentally from later abstraction and pattern recognition.

Benchmark results demonstrate the architecture's effectiveness: System2 contributes substantial gains (up to +5.20 points on PersonaMem-v2) specifically on tasks requiring cross-session inference and implicit reasoning about user evolution, while contributing less to span recall tasks. This performance profile directly validates the theoretical design—the system excels precisely where traditional retrieval-based approaches struggle. For developers building multi-turn AI agents and personalized systems, this research suggests that memory organization matters as much as memory content, and that separating recording from reasoning processes yields better long-term adaptation. The work has implications for building more sophisticated AI assistants that understand and respond to gradual user preference changes across extended interactions.

• →DCPM reorganizes agent memory into a hierarchy from raw inputs to cross-domain patterns, addressing limitations of flat retrieval-based systems
• →Dual-process architecture separates synchronous belief recording (System1) from asynchronous schema induction (System2), mimicking human memory consolidation
• →System2 improvements yield up to +5.20 point gains on implicit cross-session personalization benchmarks, validating the architectural approach
• →The system excels at reasoning over user evolution and implicit patterns while maintaining expected performance on direct recall tasks
• →Results suggest memory organization strategy fundamentally affects LLM agent capability for long-term personalization and implicit reasoning