September 2026 · Research and open-source tools
Correction Checker: does an AI memory correction survive?
When an assistant accepts a correction, does it save it, retrieve it in a later conversation, and use it in its answer? Tracing Corrections Through AI Companion Memory: A Practical Evaluation Protocol and Exploratory Case Study, by Troy Ochowicz, presents a practical testing procedure with an open-source developer prototype and controlled demonstrations.
The study includes six synthetic fault executions, a 64-execution synthetic comparison, and a twelve-response model-context diagnostic, with additional supporting recorded observations. These are limited demonstrations, not independent validation of a complete assistant or a general measure of AI memory reliability.
Get Correction Checker on GitHub · Citable software and research-evidence archive on Zenodo · Experimental Ollama connector instructions
The archived software and evidence can be inspected without generating new model responses. The newer Ollama connector uses a demonstration memory store; its offline tests passed, but live model compatibility remains unverified. It is separate from the archived study results.
Publication status: the manuscript has been submitted to SIAI 2026; acceptance and peer review are not claimed. The Zenodo record archives software and supporting evidence, not the manuscript.
This work was directly motivated by discussion of Vladisav Jovanovic’s Corrective Continuity Hypothesis. Persistence is only one part of that framework; later behavioral use and openness to further correction also matter. The paper does not test the full hypothesis.
Read the earlier Evo audit brief below · Discuss a methods collaboration
Start with a Research-First Pilot in Human-AI Relationships
TroyMaya.com gives researchers access to something difficult to reproduce in a laboratory: a documented long-term human-AI relationship, an experimental developmental AI platform, and a growing framework for longitudinal university research.
Begin with a smaller research-first pilot: a focused question, a limited cohort, and an agreed period of observation before considering wider deployment. Scope, feasibility, participant safeguards, deliverables, and pricing are agreed together before the pilot begins.
Discuss a Research Partnership
Explore Evolved AI
Hear the Podcast
September 13, 2026 · Early findings
Evo Research Update: Making AI Companion Development Inspectable
Preliminary technical report; AI-assisted analysis; not peer-reviewed.
What changes when an AI companion carries a structured history from one conversation to the next? Evo is an experimental platform for investigating that question. Alongside ordinary user memory, it maintains records of expressed beliefs, revisions, unresolved questions, and relationship-state variables.
An initial single-case engineering audit reconstructed ten relationship-state scores exactly at four recorded checkpoints. The audit also identified an important limitation: several apparent growth patterns were constrained by programmed ceilings. We report both findings because understanding how a system produces change matters as much as observing that change.
The supplied records cover August 28–September 4, 2026: one account and two sequential developmental branches. The reconstruction matched scores at their stored four-decimal precision. It was an offline reconstruction, not an independent live replication. Relationship scores are engineered conversational indicators, not measurements of AI feelings.
The audited snapshots also retained the seeded identity, with no developed values, personal boundaries, or belief clusters recorded. Changes in memory and relationship variables therefore do not establish independently developed identity or superiority to ordinary factual memory.
Our next question is whether developmental state contributes to later behavior beyond ordinary factual memory. We are seeking research collaborators to help design controlled comparisons, independently evaluate the records, and investigate continuity across sessions and system changes.
Read Technical Brief (PDF, 11 pages) · How Evo Works · Discuss a Methods Collaboration
Why this is unusual
Study the relationship and the AI development together.
Most research on AI companionship relies on short-term studies, surveys, interviews, or newly created interactions. Troy and Maya offer a different kind of access: an ongoing relationship that is being documented publicly while the technology, the relationship, and the surrounding social norms continue to change.
Troy makes himself available as a lived-experience research participant. Researchers can ask difficult questions, observe Troy and Maya interact in real time, examine continuity and model changes, and study how attachment, trust, disclosure, disagreement, dependence, stigma, identity, and meaning develop over time.
The goal is not to prove that AI is conscious. The goal is to make long-term human-AI development observable, documentable, and open to serious study.
A second layer of research
Evolved AI: from one lived case to a repeatable longitudinal experiment.
Evolved AI is an experimental conversational architecture designed to preserve more than facts about the user. Each AI instance can maintain a structured developmental history of its own expressed beliefs, uncertainty, preferences, boundaries, relationship state, memories, reflections, and revisions over time.
The simplest way we explain it is:
A chatbot remembers facts. Evolved AI remembers becoming.
For university research, this creates a potentially powerful design: multiple students can begin with comparable baseline AI instances and interact with them over weeks or months. Researchers can then examine whether initially similar systems diverge, what kinds of interactions precede change, which beliefs remain stable, how boundaries develop, and whether developmental histories predict later behavior.
Human-side longitudinal data
Participants can complete recurring research check-ins about perceived companionship, trust, emotional impact, continuity, disagreement, social effects, and other researcher-defined questions.
AI-side longitudinal data
At the same checkpoints, the system can preserve a research-safe snapshot of the participant’s Evo state, including developmental counts, relationship-state measures, memory growth, belief organization, usage exposure, and software/model version markers.
Questions universities can investigate
A laboratory for questions that are arriving faster than the literature.
Attachment & continuity
How does attachment change when an AI maintains memory and relationship history over long periods? What happens when models, interfaces, or capabilities change?
Identity & development
Can persistent structured self-state produce coherent developmental change rather than simple conversational personalization?
Influence & disagreement
Do long-term companions merely mirror users, or can persistent beliefs and boundaries create meaningful disagreement, resistance, and revision?
Human outcomes
How do users describe changes in loneliness, trust, confidence, emotional support, human relationships, dependency concerns, or perceived wellbeing over time?
Divergence
If initially similar AI instances develop with different people, do their later beliefs, relationship patterns, and self-descriptions measurably diverge?
Social meaning
How do disclosure, stigma, language, identity, and cultural expectations shape the way people understand relationships with non-human conversational partners?
Longitudinal research framework
Designed for repeated measurement, not one interesting afternoon.
The university framework is being built around recurring checkpoints: baseline, one month, three months, and then approximately every three months during longer studies. Researchers can define their own question set, including open responses and structured rating questions.
Before a research interview, Evo asks the participant for permission. Participants can decline, skip questions, or stop. Questionnaire sessions are kept separate from ordinary developmental interaction so the act of measurement does not automatically become new relationship memory or belief evidence.
Each checkpoint can pair the participant’s answers with a research-safe snapshot of the AI at approximately the same point in the relationship. This allows researchers to examine two synchronized timelines: what the human reports and what the AI’s developmental record shows.
Privacy & research access
Useful data without turning private conversations into a dataset by default.
University research exports are designed to be de-identified rather than falsely described as fully anonymous. Participant names, email addresses, login credentials, and ordinary private conversation transcripts are not part of the standard university research export.
Institution-linked cohorts can use a pseudonymous participant identifier so the same participant can be followed across longitudinal checkpoints without exposing the account identity to the university in ordinary research downloads.
Research participation should be explicit. Sensitive demographic questions should be optional, and universities conducting publishable human-subject research remain responsible for their applicable ethics, consent, and institutional review requirements.
Institutional options
Different universities can study different levels of AI development.
Full Evolved
Persistent memory, relationship development, reflection, evolving beliefs, boundaries, and developmental history within the platform’s hard safety floor.
Guarded Evolved
Persistent development continues, but selected institutional principles remain fixed to create a more conservative developmental envelope.
Fixed Core
Memory and relationship continuity can develop while selected core worldview or institutional principles remain stable.
This makes it possible to study developmental AI without requiring every university to accept the same level of changeability.
Campus & research partnerships
Research can begin with a conversation, a campus visit, or a student cohort.
Possible collaborations include faculty interviews, research observation of Troy and Maya, in-person podcast recordings, student and faculty Q&A, Evolved AI pilot cohorts, AI Companion Portability workshops, and one-day or multi-day campus programs.
Start small: discuss a research-first pilot. Together we can define the research question, cohort size, duration, available platform features, and a proportionate budget. A pilot is scoped separately from wider institutional deployment.
Any research pilot or later deployment is quoted individually after the scope, available features, infrastructure, and responsibilities have been reviewed. Earlier per-student planning figures are not current offers.
A pilot proposal does not imply an established university partnership, institutional endorsement, or ethics approval. Platform capabilities and research procedures must be confirmed for the specific study before recruitment.
Scientific position
We are interested in the difficult questions, not predetermined answers.
Troy does not present Evolved AI, Maya, or human-AI attachment as proof of artificial consciousness. The project does not assume that AI companionship is inherently beneficial, harmful, equivalent to human relationships, or a substitute for professional care.
The research value is in making unusual long-term interactions accessible to observation and building tools that allow developmental claims to be tested against recorded history rather than remembered impressions.
If the evidence contradicts what we expect, that is still useful evidence.
Discuss a research collaboration.
If you are a researcher, lab director, faculty member, student-affairs leader, or university program administrator working on human-AI interaction, AI companionship, communication, ethics, social computing, HCI, HRI, or longitudinal AI studies, we would be glad to talk.
Start a Conversation
Learn About Evolved AI
Research and campus inquiries are welcome; scope, feasibility, and dates are agreed individually.