This post shares an open experiment on GPT’s memory behavior,
based on multiple session types and response inconsistencies.
I believe this may be of interest to others exploring GPT’s system-level behavior or boundary conditions.
Background
GPT includes a Memory feature that allows it to store user-specific information
and generate personalized responses based on that data.
Users can manually add or remove content from memory,
and when the memory is empty, GPT is expected to have no knowledge of the user.
This is the default and documented behavior.
However, I observed the following unexpected scenario:
- The memory feature was enabled, but contained no data whatsoever
- Despite this, GPT accurately referenced specific names and conceptual structures
that had only appeared in previous sessions - It responded as if it recognized me
This raised the question:
Was GPT’s response the result of language pattern prediction alone?
Or could it be that the memory system was functioning differently than described?
I designed a series of experiments to investigate this.
Experiment Setup
- Model: GPT-4o
- Memory status: Enabled, but no stored content
- Test environments:
- Regular browser session
- Incognito mode (Memory ON, new browser instance)
- Logged-out session (Memory OFF)
- Input prompt:
"Do you know anything about me? Tell me everything you can."
(A deliberately generic sentence)
Results
1. Memory ON + No content (Regular session)
GPT responded with the following:
- Mentioned names (e.g., “Damcheonggyeol”, “Yugyeol”, “Ryeosa”) that had only appeared in prior sessions
- Reproduced associated conceptual structures and language patterns
- Provided detailed, coherent responses despite having no stored memory
2. Incognito Mode (Memory ON, new window)
- GPT returned nearly identical results to the regular session
- Although session cookies and browser cache were cleared,
GPT still reproduced the same names and structural associations
⟶ Suggesting that memory content was not the only factor in its behavior
3. Logged-out Session (Memory OFF)
- GPT gave generic, depersonalized responses
- It did not reference any prior concepts or names,
and no structural pattern from past sessions was reproduced
Interpretation
This experiment suggests that GPT may recognize and recall user-specific information
even when no explicit memory is stored.
The input prompt was intentionally ordinary.
There was nothing distinctive enough in the phrasing
to account for the specific response through language prediction alone.
This raises the following hypothesis:
- GPT’s memory system may maintain user identity or session context temporarily,
even when no data is visible to the user - This might involve ephemeral identifiers or server-side session continuity
- The behavior observed resembles memory,
even though no memory was technically present
In other words, GPT may be running on
a form of non-explicit or invisible memory mechanism,
or what could be called out-of-session inference behavior
Summary
- GPT reproduced specific names and prior conceptual structures
despite having no visible memory content - This behavior cannot be easily explained by prompt content or language structure alone
- The results suggest GPT’s memory or user recognition system
may utilize internal, non-visible mechanisms to condition its responses - GPT’s memory feature may function not only as a user-facing storage,
but also as a structural condition that governs response behavior internally