cognitive-mri-conversations
Research compendium: Cognitive MRI of AI Conversations. Two conference papers: Complex Networks 2025 (published, Springer) and ISCS 2026 (accepted; Springer proceedings due December 2026). A temporal-network extension is seeking a venue.
Resources & Distribution
Source Code
Package Registries
Cognitive MRI of AI Conversations: Research Compendium
Applying complex network analysis to ChatGPT conversation archives to reveal knowledge organization, community structure, and temporal evolution patterns.
Authors: Alexander Towell and John Matta, Southern Illinois University Edwardsville
Overview
This is the umbrella for a research program that transforms sequential AI conversation logs into semantic similarity networks, revealing latent cognitive structure in AI-assisted knowledge exploration. The program analyzes 1,908 ChatGPT conversations spanning December 2022 to April 2025. This repository is a landing page: it indexes the individual paper repositories and points to the shared data and pipeline.
For current priorities, existing work and future options, start with the research plans and metadata. The immediate priority is an empirical ISCS 2027 paper, with semantic/embedding dynamics retained as a second track. The dynamics consolidation assessment compares the two related drafts; their manuscripts and repositories remain separate.
Papers
Most paper projects are independent Git repositories under papers/ (not tracked by this
umbrella). The synthesis draft and data-paper manuscript live in the additional locations
listed below. Clone child repositories individually. Statuses below reflect the October 16–17,
2026 review; the full catalog
links the manuscripts and publication evidence.
| Paper | Venue / status | Path | Repository |
|---|---|---|---|
| Cognitive MRI of AI Conversations (semantic embedding networks) | Complex Networks 2025; Springer chapter published February 2026 | papers/knowledge-networks/ | cmri-knowledge-networks |
| Temporal Evolution of Cognitive Knowledge Networks | PLOS Complex Systems (rejected 2026-04-27, seeking venue) | papers/temporal-networks/ | cmri-temporal-networks |
| From Episodes to Abstractions: Latent Hierarchical Memory | ISCS 2026 accepted; full proceedings forthcoming | papers/hierarchical-memory/ | cmri-hierarchical-memory |
| Embedding Dynamics | Partial draft; proposed consolidation with Semantic Dynamics | papers/embedding-dynamics/ | cmri-embedding-dynamics |
| Semantic Dynamics | Draft; proposed consolidation with Embedding Dynamics | papers/semantic-dynamics/ | cmri-semantic-dynamics |
| Multi-Scale Network Organization in AI-Assisted Knowledge Exploration | Unsubmitted synthesis draft | future-ideas/compsac-2026/paper/ | This umbrella repository |
| Sanitized ChatGPT Conversation Corpus | Unpublished data-paper draft; release status unresolved | chatgpt-conversation-corpus/paper/ | chatgpt-conversation-corpus |
| Operational Memex | Outline and working infrastructure | papers/operational-memex/ | cmri-operational-memex |
Pilot studies (October 2026)
Nine bounded pilots, mostly on public multi-participant data, each in its own private
repository under papers/ with a REPORT.md and claim ledger:
cmri-portfolio-wave1,
cmri-history-falsification,
cmri-history-specificity,
cmri-grounded-dependencies,
cmri-integration-measurement,
cmri-semantic-measurement,
cmri-selective-forgetting, plus the
temporal calibration on the pilot/iscs-2027-calibration branch of
cmri-temporal-networks. The
decision history records what each one established.
Confirmatory studies followed the pilots, each a frozen single-shot run on samples no analysis had touched. The primary contrast (uniform history pooling against the last two messages, next user message) passed in both encoders on the held-out PRISM reserve (cmri-iscs-2027) and on English WildChat-4.8M (cmri-wildchat-replication). A second protocol then confirmed what that gain is, the person’s own preceding turn, on fresh PRISM and WildChat samples, the latter drawn by cmri-history-reach. No held-out sample may be reused for further confirmation.
The published conference paper and the shared analysis pipeline (code/ and data/) live
in the knowledge-networks repository; the temporal-evolution journal extension is
temporal-networks.
How this repository is organized
papers/: the independent paper, pilot and study repositories (gitignored here).workspace.jsonlists every one with its path, branch, visibility and data class, andpython3 tools/workspace.py cloneclones them into place (status,verifyand the other subcommands are described intools/workspace.py).chatgpt-conversation-corpus/: the sanitized, publishable conversation dataset, its own git repository (gitignored here). Derived fromdev/; pre-release, sanitization in progress.dev/: shared working corpus (gitignored; the raw conversation JSON the pipelines read).future-ideas/: early-stage directions not yet promoted to their own repos.docs/: program-level design and planning documents.
Shared data and pipeline
- Conversation corpus (dataset): chatgpt-conversation-corpus, co-located in this cluster at
chatgpt-conversation-corpus/. The sanitized, publishable version of the corpus (pre-release; sanitization in progress). - Analysis pipeline of record: chatgpt-complex-net (DOI: 10.5281/zenodo.15314235)
Child code reads the raw corpus from $CMRI_CORPUS_DIR, which defaults to this repo’s dev/.
Citation
Cite the specific paper you use (each paper repository has its own CITATION.cff). To cite
the program as a whole, use the compendium DOI:
10.5281/zenodo.18778376. See
CITATION.cff for machine-readable metadata.