Follow memory from a relevant moment in a conversation to what the assistant says, what the user signals, and whether they return.
We compare people with similar prior activity and similar memory opportunities. This helps separate a memory signal from the simple fact that highly engaged people tend to return.
This is a synthetic model of a ChatGPT-style assistant. It does not measure ChatGPT’s internal memory or real users. In the default scenario, retrieval has zero simulated effect on return.
The controls filter the cohort snapshot, example, and return views. Segment cards and the opportunity catalog stay on the full cohort; pending follow-up and unknown telemetry remain visible.
How often memory was retrieved at an eligible moment.
Whether retrieved items were relevant, current, and used appropriately.
Whether users came back later. The pattern can reflect prior engagement or other causes, so it is not proof memory caused a return.
Each card shows the full synthetic cohort, even when an opportunity filter is selected. A user can be eligible for multiple segments.
Episodic, semantic, and procedural memory are long-term stores. Retrieval brings information from those stores into working memory; it does not mean the assistant used it well.
These example rules are set before retrieval or return is examined. The table shows the full synthetic cohort as a reference catalog.
This example makes the event labels concrete. It illustrates the measurement; it is not evidence about actual users.
Review what was returned, whether it reached the response, and what the user explicitly signaled.
Compare raw return rates, then standardize across prior engagement and opportunity type. Return is an activity measure, not a memory-quality score.
It can demonstrate an analysis framework and teach the limits of observational retention comparisons. It cannot establish ChatGPT product findings.
The assistant interacts with the user or another external environment. We measure observable response support and user feedback.
The system reads long-term memory into the active workspace. We record segment, result status, relevance, and freshness.
The assistant combines current and retrieved information. We evaluate the visible answer; we do not collect private chain-of-thought.
The system writes, updates, or deletes long-term memory. We check correction persistence and forget compliance.
Portfolio disclosure: all records and example conversations are synthetic. The site does not connect to ChatGPT, any company system, database, or external data service.