Here's what's actually happening under the hood—and why you should be concerned.

How the Dreaming V3 Memory Engine Works

Dreaming V3 abandons the explicit saved-memories list entirely. Instead, a background synthesis engine runs after conversations end, reading across a user's full chat history to infer preferences, work patterns, personal constraints, and other details—all without any manual "remember this" command .

Key architectural changes include:

• Full-history scanning. The system no longer relies on items you've flagged—it reads everything .
• Inferred preferences. It extracts implicit details from casual remarks and past behavior that you may never have intended to store .
• Between-session consolidation. A dreaming process runs offline, resolving contradictions and updating stale facts .
• Synthesized representation. What gets stored isn't a verbatim transcript of your words but a compressed, model-generated description of who you are .

Users can view a memory summary in Settings, but the underlying raw data layer is not directly exposed . That creates an audit gap that becomes critical when the system starts getting things wrong.

ZDNet's Testing Found Widespread Accuracy Failures

ZDNet reporter David Gewirtz tested the upgraded memory extensively and found serious, systematic accuracy problems that compound over time .

• Outdated assumptions stick permanently. ChatGPT stored old job titles, expired projects, and superseded preferences, then treated them as current facts in every subsequent interaction .
• Minor errors become gospel. The AI locks in small mistakes and applies the same confidence it uses for verified factual knowledge. A single incorrect detail then silently pollutes dozens of future responses .
• Invisible propagation. Flawed memory data distorts answers without any visible indication that something is wrong. Users may never realize the AI is working from corrupted context until the damage has already spread across many conversations .
• Detailed but incorrect profiles. The system created elaborate personal profiles that were plausible but wrong, anchoring every response to false premises .

Gewirtz described Dreaming V3 as a "technical triumph" paired with an "irresponsible feature" . The core problem isn't that the AI makes errors—it's that the error becomes permanent, invisible infrastructure for everything you ask afterward.

The Audit Gap: What OpenAI Won't Show You

OpenAI's Memory FAQ contains a quietly devastating admission: "While the memory summary should capture the most important details, it will not include everything that ChatGPT remembers based on your chats. If you want to know if ChatGPT has remembered something, just ask in chat" .

The memory summary page is a human-readable overview of the synthesized profile, not a complete window into what the model actually stores and uses . That means users cannot fully audit what ChatGPT "knows" about them through the settings UI alone . You'd have to go fishing—asking the AI to reveal its own hidden assumptions about you—which is neither reliable nor practical for regular use.

Why "Just Delete It" Doesn't Fix the Problem

Options to manage or purge memory exist, but each carries significant limitations.

• Turning memory off may not erase what's already stored. ZDNet reported that disabling the feature doesn't necessarily delete the data the system has already synthesized .
• Manual deletion is fragile. Users can review and delete individual memory entries in Settings, but the dreaming process can re-learn the same details from your existing chat history the next time it runs .
• Ongoing automatic re-synthesis. Because Dreaming V3 continuously re-synthesizes context, simply deleting listed memories doesn't prevent the system from re-inferring similar information from past conversations .
• Most users won't audit until it breaks. The first visible sign of corrupted memory often comes after the poisoned data has already shaped many responses .

The New Context Rot: Rotting From the Profile Outward

Context rot describes how large language models degrade as context windows fill up—older information gets displaced by newer content, and outputs drift from the user's actual situation . Dreaming V3 was explicitly designed to combat this problem across sessions, tackling "the staleness, correctness, and scalability challenges that we observe when memory is applied to the hundreds of millions of users and multi-year time horizons in ChatGPT" .

But what OpenAI has built introduces a more insidious variant. Instead of context rotting from a finite window filling up, it now rots from the center outward: once an error enters the persistent user profile, it silently degrades every answer that follows. This isn't traditional context rot from lost information—it's rot from wrong information that the system treats as authoritative. And because users can't see or easily remove it, they may never know why their AI is slowly getting worse .

Here's what's actually happening under the hood—and why you should be concerned.

How the Dreaming V3 Memory Engine Works

Dreaming V3 abandons the explicit saved-memories list entirely. Instead, a background synthesis engine runs after conversations end, reading across a user's full chat history to infer preferences, work patterns, personal constraints, and other details—all without any manual "remember this" command .

Key architectural changes include:

• Full-history scanning. The system no longer relies on items you've flagged—it reads everything .
• Inferred preferences. It extracts implicit details from casual remarks and past behavior that you may never have intended to store .
• Between-session consolidation. A dreaming process runs offline, resolving contradictions and updating stale facts .
• Synthesized representation. What gets stored isn't a verbatim transcript of your words but a compressed, model-generated description of who you are .

Users can view a memory summary in Settings, but the underlying raw data layer is not directly exposed . That creates an audit gap that becomes critical when the system starts getting things wrong.

ZDNet's Testing Found Widespread Accuracy Failures

ZDNet reporter David Gewirtz tested the upgraded memory extensively and found serious, systematic accuracy problems that compound over time .

• Outdated assumptions stick permanently. ChatGPT stored old job titles, expired projects, and superseded preferences, then treated them as current facts in every subsequent interaction .
• Minor errors become gospel. The AI locks in small mistakes and applies the same confidence it uses for verified factual knowledge. A single incorrect detail then silently pollutes dozens of future responses .
• Invisible propagation. Flawed memory data distorts answers without any visible indication that something is wrong. Users may never realize the AI is working from corrupted context until the damage has already spread across many conversations .
• Detailed but incorrect profiles. The system created elaborate personal profiles that were plausible but wrong, anchoring every response to false premises .

Gewirtz described Dreaming V3 as a "technical triumph" paired with an "irresponsible feature" . The core problem isn't that the AI makes errors—it's that the error becomes permanent, invisible infrastructure for everything you ask afterward.

The Audit Gap: What OpenAI Won't Show You

OpenAI's Memory FAQ contains a quietly devastating admission: "While the memory summary should capture the most important details, it will not include everything that ChatGPT remembers based on your chats. If you want to know if ChatGPT has remembered something, just ask in chat" .

The memory summary page is a human-readable overview of the synthesized profile, not a complete window into what the model actually stores and uses . That means users cannot fully audit what ChatGPT "knows" about them through the settings UI alone . You'd have to go fishing—asking the AI to reveal its own hidden assumptions about you—which is neither reliable nor practical for regular use.

Why "Just Delete It" Doesn't Fix the Problem

Options to manage or purge memory exist, but each carries significant limitations.

• Turning memory off may not erase what's already stored. ZDNet reported that disabling the feature doesn't necessarily delete the data the system has already synthesized .
• Manual deletion is fragile. Users can review and delete individual memory entries in Settings, but the dreaming process can re-learn the same details from your existing chat history the next time it runs .
• Ongoing automatic re-synthesis. Because Dreaming V3 continuously re-synthesizes context, simply deleting listed memories doesn't prevent the system from re-inferring similar information from past conversations .
• Most users won't audit until it breaks. The first visible sign of corrupted memory often comes after the poisoned data has already shaped many responses .

The New Context Rot: Rotting From the Profile Outward

Context rot describes how large language models degrade as context windows fill up—older information gets displaced by newer content, and outputs drift from the user's actual situation . Dreaming V3 was explicitly designed to combat this problem across sessions, tackling "the staleness, correctness, and scalability challenges that we observe when memory is applied to the hundreds of millions of users and multi-year time horizons in ChatGPT" .

But what OpenAI has built introduces a more insidious variant. Instead of context rotting from a finite window filling up, it now rots from the center outward: once an error enters the persistent user profile, it silently degrades every answer that follows. This isn't traditional context rot from lost information—it's rot from wrong information that the system treats as authoritative. And because users can't see or easily remove it, they may never know why their AI is slowly getting worse .