From Episodes to Abstractions: Latent Hierarchical Memory in 1,908 AI Conversations

Published on March 30, 2026 Accepted

Authors:
Alexander Towell (lex@metafunctor.com)
John Matta (jmatta@siue.edu)

Abstract

We investigate whether AI conversation archives contain hierarchical structure analogous to human semantic memory. From a longitudinal archive of 1,908 ChatGPT conversations spanning 29 months, we extract semantic concepts using LLM-based analysis, embed them in a 768-dimensional semantic space, and apply hierarchical agglomerative clustering to organize them into a four-level hierarchy: 500 meta-concepts, 50 themes, and 8 knowledge domains.

Download PDF

Cite This Work

Show BibTeX
#semantic memory #hierarchical clustering #AI conversations #complex networks #small-world networks