CogniFold: Always-On Proactive Memory via Cognitive Folding

arXiv:2605.13438 · cs.AI, cs.CL · Submitted 2026-08-22 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "CogniFold: Always-On Proactive Memory via Cognitive Folding".

Jane: The paper was written by Suli Wang, Dai Shi, Yiqun Duan, Minghua Deng, Yu Deng et al. from University of Cambridge and OpenNorve and NVIDIA and Griffith University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: So, given the title, CogniFold: Always-On Proactive Memory via Cognitive Folding, how do you even begin to imagine what "proactive" means in an AI agent?

Jane: It’s about having the assistant surface information before we even ask for it, which is a big leap from simply waiting for us to speak.

Lu: The idea of "always-on" suggests that this isn't a system that wakes up and processes one batch of data; it’s constantly processing the stream as it flows in.

Meng: That constant flow is the major engineering challenge, because traditional systems just can't handle that kind of endless, asynchronous input.

Lalam: It means we are designing agents that are already anticipating our needs, creating a truly anticipatory relationship with the user.

Summary: Tom: The core concept is this tri-layered architecture—Hippocampus, Neocortex, and Prefrontal Intent Layer—how does that work in simple terms?

Jane: Think of it like three different stages of processing; the Hippocampus captures raw episodes, the Neocortex abstracts those into general concepts, and then the Prefrontal layer figures out what all those concepts mean for future goals.

Lu: It’s a biological model translated into a typed multigraph that constantly metabolizes information rather than just a fixed structure.

Meng: The key is that it' continuous folding loop allows the system to learn while it's operating, which is something most existing architectures simply don't do at scale.

Lalam: This means the AI isn't just recalling facts; it’s synthesizing a deeper understanding of our habits and routines from accumulating data streams.

Improvements: Tom: The authors highlight that this system addresses four specific "structural debts" that continuous input creates, which are accumulation, compression, decay, and completion.

Jane: That is a sophisticated way of saying the memory naturally fixes its own structural problems as the data flows in.

Lu: It’s not just patching things up; it’s built into the very topology of merging and reinforcing concepts that is fascinating from a theoretical perspective.

Meng: From an implementation standpoint, addressing those four debts automatically means we' don't need separate post-processing steps to clean up the memory.

Lalam: The idea of "cognitive folding" allows for this massive compression while ensuring the memories remain highly relevant and grounded in our actual experiences.

Conclusion: Tom: So, looking at the results, we see that CogniFold not only handles these technical demands but also generates a level of proactive intent that was previously unachievable.

Jane: The ability to generate those intent nodes means the system is truly self-organizing rather than just pulling information from a list.

Lu: It provides real evidence of cognitive bootstrapping, where the structure is actively building itself up based on its own past experiences.

Meng: The four point six times compression and that zero point six one four proactivity rate shows we've found a way to make these systems incredibly efficient and useful in practice.

Lalam: It’s not just about better memory; it’s about creating an AI that has a genuine, living understanding of our lives, which is the ultimate goal for AI assistants.

Tom: That is a powerful idea for an always-on agent. Thank you all so much for sharing your insights into CogniFold: Always-On Proactive Memory via Cognitive Folding with us today.

Lu: I’m excited to see how this architecture can be used in more advanced cognitive modeling, Tom.

Meng: I think this will dramatically change how we design memory in our own AI products, making it much more practical for real-world use cases.

Lalam: It offers a pathway toward truly symbiotic interaction between the user and the machine.

Jane: It’s clear that this work is moving us toward a genuinely proactive relationship with AI.

Suli Wang, Dai Shi, Yiqun Duan, Minghua Deng, Yu Deng, Chen Chen, Rundong Zhao, Yiqi Wang

University of Cambridge · OpenNorve · NVIDIA · Griffith University

cs.AI, cs.CL

Submitted: 2026-08-22

Updated: 2026-08-25

Code: https://github.com/OpenNorve/CogniFold

Importance score: 85/100

The gist: CogniFold: Always-On Proactive Memory via Cognitive Folding The paper addresses a fundamental limitation in existing agent memory architectures, which are described as "predominantly reactive and

Key concepts

Proactive Memory
This concept allows an AI assistant to surface information before the user even asks for it. Unlike traditional systems that wait for input, this approach creates a truly anticipatory relationship with the user by constantly processing data streams to meet anticipated needs.
Tri-layered Architecture
The system uses three distinct layers: The Hippocampus captures raw episodes, the Neocortex abstracts these into general concepts, and the Prefrontal Intent Layer determines what those concepts mean for future goals. This biological model allows for continuous information metabolism.
Cognitive Folding
This refers to a continuous folding loop that enables the AI to learn while it is operating. It provides a mechanism for massive data compression while ensuring that memories remain highly relevant and grounded in the user's actual experiences.

Terminology

Summary

CogniFold: Always-On Proactive Memory via Cognitive Folding

The paper addresses a fundamental limitation in existing agent memory architectures, which are described as predominantly reactive and retrieval-based, lacking the capacity to autonomously organize experience into persistent cognitive structure. Current memory systems remain a graph-as-product—a finished artifact to retrieve from, never a substrate that metabolises under the stream, forcing agents to graft proactivity onto application layers.

To move toward genuinely autonomous assistants, the authors introduce CogniFold, an always-on agent memory designed for proactive assistants. CogniFold is built on an extended form of Complementary Learning Systems (CLS) theory, which the paper formalizes as a typed, dynamically evolving multigraph. The architecture is based on a tri-layered cognitive structure:

  1. Hippocampal Layer: Event nodes serve this role, where each input from the stream is committed verbatim and time-stamped—an immutable episodic trace.

  2. Neocortical Layer: Concept nodes consolidate patterns, abstract[ing] them into schemata, anchored to their constituent events through provenance edges.

  3. Prefrontal Layer: Intent nodes emerge when evidence converges, exert[ing] top-down influence on how subsequent events are surfaced and encoded.

The core operational dynamic of CogniFold is conceptual bootstrapping, which unfolds through three continuous folding stages:

  • Stage 1: Accumulation. The system ingests the raw stream verbatim into Event nodes.

  • Stage 2: Consolidation. Statistical regularities across events are detected, and discrete Event nodes are folded into Concept nodes anchored to their grounding events.

  • Stage 3: Crystallization. Concepts act as scaffolds for future input; when concept-cluster density crosses a threshold, an Intent node crystallizes in the prefrontal layer, providing top-down bias.

This continuous folding process is designed to manage four intrinsic structural debts that any constantly evolving graph must accrue:

  1. Accumulation: The system ensures persistent patterns must strengthen; one-off noise must not by creating a REINFORCES edge when a new event corroborates an existing concept.

  2. Compression: Redundant fragments are merged via the MERGE NODES operation when two concept nodes exceed a semantic-similarity threshold.

  3. Decay: Stale structure weakens, as all edges undergo exponential decay at every consolidation pass.

  4. Completion: Missing connections are inferred through kNN inference over concept embeddings, which "scans for zero-edge concept nodes and creates GROUND connections—automatically repairing gaps the LLM’s local-view planning misses."

The system utilizes a Proactive Context Assembly mechanism on the write path. Priority for which knowledge is presented to the LLM is determined by a weighted score combining structural centrality (Personalized PageRank), temporal recency, and access intensity:

Score(v) = alpha times PR(v) + beta times e(-lambda times t v) + gamma times Acc(v) times U(v)

This allows the system to surface relevant context before the next event is interpreted—rather than waiting for a later query to reveal what should have mattered.

Evaluation and Results:

The authors validate CogniFold using two methods:

  1. Structural Evaluation (CogEval-Bench): This framework measures whether the topology formed under continuous event streams matches cognitive expectations, demonstrating that CogniFold uniquely produces event-grounded concepts, coherent conceptual structure, and proactive intent emergence.

  2. Downstream Utility: The system is tested across eight benchmarks—two long-term conversational memory tests (LoCoMo and LongMemEval) and six other cognitive domains.

The results demonstrate significant superiority:

  • In CogEval-Bench, C OGNI F OLD was the only system producing non-zero purity and proactivity.

  • The system achieved 4.6× compression of events into concepts.

  • In long-term conversational memory (LongMemEval), CogniFold led by more than twenty points in the critical 'Build' column (93.0% overall).

  • Across the six cognitive domains, C OGNI F OLD consistently outperformed baselines, demonstrating that cognitive folding—the operations of §3.2—is a task-general write-path competence rather than a benchmark-tuned heuristic.

Improvements for AI systems

The current work presents a sophisticated cognitive memory architecture (C OGNI F OLD) that successfully moves beyond simple event extraction towards structured, integrated concept emergence and proactive intent identification. However, given the high stakes and need for industrial robustness, several critical areas require methodological and architectural improvements.

Here are the specific improvements I recommend for the AI system:

Current Limitation: The system generates timestamped events, which establishes when things happened. However, the integration between concepts often remains purely associative or co-located in time. It lacks a formal model for causality or necessary preconditions.

Improvement: Integrate a Probabilistic Causal Inference Module (PCIM) into the graph construction pipeline.

  • Mechanism: When linking two events (e A and e B) or linking an event to a concept node (c), the PCIM must calculate the probability of causality (e.g., P(BA)) rather than just measuring co-occurrence or similarity. This involves training a structured causal model (like a Bayesian Network or Granger Causality test applied to textual features) on the LLM's generated context.

  • Output: The graph edges must be augmented with a Causality Score (C) and an Inference Type Label (e.g., Prerequisite, Stimulus, Consequence).

  • Improved Capability: The system can move from merely stating that Concept X was active during the time of Event Y to definitively stating: "The realization of Concept X at time t A was a necessary prerequisite for the subsequent event e B at time t B, with a causal probability C=0.92." This drastically improves predictive accuracy and forensic analysis.

By implementing these three modules—PCIM (Causality), MCCFL (Modality), and CSE (Counterfactual)—the resulting cognitive memory system transforms from a highly advanced Associative Knowledge Graph into a Dynamic, Predictive, Causal Simulation Engine. It moves from answering What happened? to answering the far more valuable questions: "Why did it happen in this sequence? and What must happen next for the desired outcome to be achieved?" This leap in capability is essential for deployment in high-stakes fields like personalized medicine, complex operational decision support, and advanced robotics.

Sources

Related papers