Towards LifeSpan Cognitive Systems

TL;DR

The paper defines LifeSpan Cognitive System (LSCS) as a long-lived, human-like agent that continuously learns from rich, high-frequency experiences and must solve two core challenges: Abstraction and Experiences Merging, and Long-Term Retention with Accurate Recall. It introduces a Storage Complexity taxonomy that partitions past-experience storage into four classes: Saving $\mathcal{E}$ into Model Parameters ($0$), Saving $\mathcal{E}$ into Explicit Memory ($o(n)$ or $O(n)$), Saving $\mathcal{E}$ into Knowledge Bases for Retrieval, and Saving $\mathcal{E}$ into Raw Text and Process All, and then analyzes the benefits and limitations of each. The authors propose a conceptual LSCS instantiation that combines these approaches into two phases: Absorbing Experiences and Generating Responses, leveraging model parameters, explicit memory, and knowledge bases to achieve both abstraction and robust recall. This framework provides a roadmap for building lifespan-scale cognitive systems capable of rapid adaptation and coherent long-horizon reasoning in dynamic environments, though practical implementation and benchmarking remain open challenges. The discussion emphasizes integrating memory tokens, RAG, and long-context processing to realize a practical, scalable LSCS.

Abstract

Building a human-like system that continuously interacts with complex environments -- whether simulated digital worlds or human society -- presents several key challenges. Central to this is enabling continuous, high-frequency interactions, where the interactions are termed experiences. We refer to this envisioned system as the LifeSpan Cognitive System (LSCS). A critical feature of LSCS is its ability to engage in incremental and rapid updates while retaining and accurately recalling past experiences. In this paper we focus on the domain of Large Language Models (LLMs), where we identify two major challenges: (1) Abstraction and Experience Merging, and (2) Long-term Retention with Accurate Recall. These properties are essential for storing new experiences, organizing past experiences, and responding to the environment in ways that leverage relevant historical data. Unlike language models with continual learning, which typically rely on large corpora for fine-tuning and focus on improving performance within specific domains or tasks, LSCS must rapidly and incrementally update with new information from its environment at a high frequency. Existing technologies with the potential of solving the above two major challenges can be classified into four classes based on a conceptual metric called Storage Complexity, which measures the relative space required to store past experiences. Each of these four classes of technologies has its own strengths and limitations while we argue none of them alone can achieve LSCS alone. To this end, we propose a potential instantiation for LSCS that can integrate all four classes of technologies. The new instantiation, serving as a conjecture, operates through two core processes: Absorbing Experiences and Generating Responses.

Figures (2)

• Figure 1: The technologies for constructing Lifespan Cognitive Systems (LSCS) can be broadly categorized into two principal approaches based on storage complexity. The first approach, characterized by zero storage complexity, involves Saving $\mathcal{E}$ into Model Parameters. The second approach, which involves non-zero storage complexity, is subdivided into methods with storage complexities of $o(n)$ and $O(n)$. The $o(n)$ methods indicate Saving $\mathcal{E}$ into Explicit Memory. The $O(n)$ methods are further classified based on whether the language model in the system processes the entire stored text. Suppose an additional retriever is adopted and the language model only accesses a snippet of the stored experiences. In that case, it falls under the category of Saving $\mathcal{E}$ into Knowledge Bases for Retrieval. The final category encompasses methods where the entire context is input into the model, classified as Saving $\mathcal{E}$ into Raw Text and Process All.
• Figure 2: The Operating Diagram of LSCS mainly includes two parts: (a) Abstraction and Experiences Merging; (b) Generating responses according to the environmental query, where the ability of long-term retention and accurate recalling should be guaranteed. Note that we add "Notepad" after "Knowledge Base", which is meant as an analogy. See Section \ref{['sub:absorbing_experiences']} for more details.