Current AI agents rely on reactive, retrieval-based memory that cannot autonomously structure experience into persistent cognitive formations. This limitation hinders the development of truly autonomous agents capable of proactive reasoning and long-term adaptation. Addressing this gap, a new research paper introduces CogniFold, a brain-inspired 'always-on' agent memory designed for next-generation proactive assistants.
CogniFold continuously folds fragmented event streams into self-emerging cognitive structures, bootstrapping progressively higher-level cognition from incoming events and accumulated knowledge.
Memory Architecture: Three-Layer Cognitive Folding
The researchers extend Complementary Learning Systems (CLS) theory from two layers — the hippocampus and neocortex — to three, adding a prefrontal intent layer. This layer emulates the prefrontal cortex as the locus of intentional control and decision-making. CogniFold achieves this through graph-topology self-organization: cognitive structures proactively assemble under the stream, merge when semantically similar, decay when stale, relink through associative recall, and surface intents when concept-cluster density crosses a threshold.
The system is designed to be always-on, meaning it continuously processes and organizes information without requiring explicit retrieval triggers. This contrasts sharply with traditional memory modules that only activate upon query.
| Traditional Reactive Memory | CogniFold Proactive Memory |
|---|---|
| Retrieval-based, reactive | Always-on, proactive |
| No autonomous organization | Self-organizing cognitive structures |
| Two-layer (hippocampus + neocortex) | Three-layer (+ prefrontal intent) |
| Static memory decay | Dynamic decay and relinking |
| No intent surfacing | Intent surfacing via density threshold |
Evaluation on Cognitive Benchmarks
The researchers evaluated structural formation using CogEval-Bench, demonstrating that CogniFold uniquely produces memory structures that match cognitive expectations and exhibit concept emergence. For downstream tasks, they tested on eight benchmarks, including two probing long-term conversational memory: LoCoMo and LongMemEval. Across all benchmarks spanning various cognitive domains, CogniFold performed robustly on conventional memory tasks.
Implications for Enterprise AI
For enterprise technology leaders, autonomous agents with persistent, self-organizing memory could transform workflow automation, customer interaction, and decision support. Systems that remember context proactively, merge related information, and surface emerging intents offer a path toward more intelligent and independent software assistants. While CogniFold is currently a research project — with code available at its GitHub repository — its architecture points to a new direction for agent memory that goes beyond simple retrieval.
The authors — Suli Wang, Yiqun Duan, Yu Deng, Rundong Zhao, Dai Shi, Minghua Chen, and Xinliang Zhou — have open-sourced the implementation, inviting validation and extension by the AI community. As the demand for autonomous systems grows in areas like supply chain management and logistics, memory architectures that emulate human cognitive organization may become a critical component.
