Guo et al. (2026)

The Interference Gap: Comparing Retrieval Bounds in Human Memory and RAG Systems

arXiv

Find It

DOI

Abstract

How do retrieval bounds compare between human episodic memory and Retrieval-Augmented Generation (RAG) systems under semantic interference? We present a unified signal detection theory (SDT) framework that applies to both, and use it to fit behavioral and computational data in matched paradigms. Both systems show logarithmic accuracy decline with association count (fan), but humans exhibit lower interference sensitivity ($α/σ= 0.41$) than dense passage retrieval ($α/σ= 0.67$), with cognitively-inspired HippoRAG falling between the two ($α/σ= 0.44$). Behavioral experiments ($N = 112$) and simulations validate the framework; parameter recovery confirms identifiability ($r \geq .93$) and model comparison favors the logarithmic specification over a power-law alternative ($Δ$BIC $> 15$). We discuss encoding specificity, temporal context binding, and retrieval gating as candidate mechanisms whose causal role remains to be established. Six falsifiable predictions connect cognitive memory research with AI retrieval evaluation.

Reference

Guo, D., Wu, J., & Yiu, S.-M. (2026). The interference gap: Comparing retrieval bounds in human memory and RAG systems (arXiv:2606.28327v1). arXiv. https://doi.org/10.48550/arXiv.2606.28327

What Rests on This

2 mechanisms cite this source. Each finding below is quoted as it appears on that mechanism's own page.

Forgetting Memory Processes

Measured in AI

The first-pass rating rested on a 2026 CogSci study of the fan effect in 112 people and retrieval systems; that measures interference between competing memories, a separate mechanism, so it is not used here

Retrieval Interference Memory Processes

Measured in humans

In a 2026 study of 112 online adults who studied 60 person-location sentences for 5 s each, then did a 2-minute arithmetic task and an old/new recognition test, accuracy fell from 91.2% at fan 1 to 82.4% at fan 2 and 71.8% at fan 4 (Cohen's d = 1.12)
Measured in AI

Deciding study: a 2026 paper (accepted at CogSci 2026 per its arXiv listing) gave three retrieval systems, the programs that look up stored text for an AI model, the same person-location sentences people studied: DPR (2020), ColBERTv2 (2022) and the hippocampus-inspired HippoRAG (2024), scored correct when the studied sentence came back first, at fans of 1, 2, 4 and 8 (only its fan and competitor results bear on this mechanism; its serial-position results bear on a different one)
Measured in AI

On a fitted interference slope (how fast discrimination falls as fan grows, on a log scale; lower is better), people scored 0.41 (95% CI 0.35 to 0.47), HippoRAG 0.44 (0.37 to 0.51), ColBERTv2 0.53 (0.45 to 0.61) and DPR 0.67 (0.58 to 0.76), so systems spanned 0.44 to 0.67 and only DPR differed reliably from people; in raw accuracy DPR fell from 94% to 58% by fan 4 while HippoRAG tracked the human curve (about 84% and 72% at fans 2 and 4)
Measured in AI

This comparison is approximate and favoured the AI: the systems held every sentence verbatim and were queried with no delay, while people had to remember after a distractor, and people judged old versus new while systems ranked stored sentences; HippoRAG matching people despite that advantage is a tie, not a win
Measured in AI

The Guo et al. numbers need independent checking: the reported slopes are about half what the paper's own human d' values (2.84, 2.21, 1.67) imply, and a query of a person's name alone, as described, could not yield the above-chance retrieval scores reported, so what the systems were asked is unclear; the study was not preregistered