AI agents are about to get a memory system that doesn't wait to be asked.
Researchers have published CogniFold, an always-on memory architecture that departs from the standard retrieval model — where agents fetch relevant past context only when prompted. Instead, CogniFold continuously processes incoming events, clustering them into evolving knowledge graphs that merge, decay, and reassemble over time. The system adds a third layer to Complementary Learning Systems theory, a neuroscience framework that previously mapped AI memory onto two brain regions: the hippocampus (fast, episodic) and the neocortex (slow, generalized). The new prefrontal layer is meant to emulate intentional control — surfacing goals and intentions autonomously when related concept clusters hit a density threshold.
Most production AI memory today is a glorified search index: you ask, it fetches. CogniFold's bet is that proactive memory — the kind that says "you probably need this" before you type — requires structure that self-organizes in the background. The researchers back this with evaluations across eight benchmarks, including two focused specifically on long-term conversational memory, where CogniFold reportedly holds its own against conventional approaches while also doing the new proactive work.
The code is open-source on GitHub, which makes the claims checkable — a low bar that plenty of AI memory papers never clear. Whether graph-topology self-organization at inference scale is practical outside a research setting is a different question entirely.