Forgetting as a Feature: Cognitive Alignment of Large Language Models

TL;DR

The paper reframes forgetting in large language models as a functional cognitive mechanism, modeling inference as a cognitively-aligned Bayesian updating process with exponential memory decay. It introduces a benchmarking suite to quantify forgetting dynamics across temporal recall, concept drift, and associative memory, and proposes Probabilistic Memory Prompting (PMP) to shape evidence integration toward human-like forgetting, with gamma controlling memory decay via $p_t(\theta|D_{1:t}) \propto p(D_t|\theta)[p_{t-1}(\theta|D_{1:t-1})]^\gamma$. Empirical results on long-horizon reasoning tasks, including a Reflective Confidence framework, show that forgetting can improve robustness and error correction, aligning model behavior with human memory dynamics. The work offers a principled path to adaptively align LLM inference with cognitive principles, enabling better performance under non-stationary contexts and long-context dependencies.

Abstract

Large Language Models (LLMs) are often evaluated against ideals of perfect Bayesian inference, yet growing evidence suggests that their in-context reasoning exhibits systematic forgetting of past information. Rather than viewing this behavior as a limitation, we reinterpret forgetting as a functional cognitive mechanism. Drawing inspiration from human memory dynamics, we model LLM inference as a probabilistic memory process governed by exponential decay. We introduce a benchmark suite that evaluates temporal reasoning, concept drift adaptation, and associative recall, enabling direct comparison between model behavior and human cognitive patterns. Our empirical results reveal that LLMs demonstrate forgetting rates analogous to human memory efficiency trade-offs between stability and adaptability. Building on these observations, we propose probabilistic memory prompting, a lightweight strategy that shapes evidence integration to mimic human-like memory decay, leading to improved long-horizon reasoning performance. Our findings position forgetting not as a failure mode, but as a principled mechanism for adaptive intelligence.

Figures (3)

• Figure 1: Overview of the Probabilistic Memory Prompting (PMP) framework. The model applies an exponential forgetting mechanism to weight historical context, samples informative memory items accordingly, and performs reasoning over the selected subset.
• Figure 2: Soft memory weighting in probabilistic prompting.
• Figure 3: Impact of the reflection threshold percentile ($p$) on GSM8K Accuracy. The setting used in our main experiments is highlighted.