Memory made ChatGPT remember you. It didn't make it present. ChatGPT's persistent memory features let the assistant retain preferences and prior exchanges across sessions when users enable them, but commentators on Substack argue that memory alone does not make a system feel like a social partner. The missing pieces include relational knowledge, cross-chat situational awareness, and the ability to model a user's beliefs, gaps that benchmarks and design proposals identify as immediate engineering priorities.

That change let assistants carry personal preferences and earlier goals from one session to the next, and it reshaped what users expect from conversational AI.

What memory actually delivered

OpenAI's memory feature moves ChatGPT out of the purely stateless prompt-response model. With the feature enabled, the assistant can preserve stylistic and factual preferences, recall constraints or long-running projects, and allow users to pick up a conversation where they left off. Practically, that means fewer repeated setup prompts and a smoother handoff across sessions when the same user returns.

Engineers and product teams describe the change as necessary. A Substack post cataloguing the development called long-term memory necessary but not enough for presence. The Substack commentary sketched a systems-level concept it named Presence Engineering, which pairs memory with continuity mechanisms, feedback loops, and situational orientation so assistants evolve with users rather than reset after each chat.

Those proposals are timely because designers are seeing obvious, fixable failure modes. A social-media account documenting daily interactions with an assistant logged a recurring annoyance: the assistant fails to notice that a related, recent conversation exists in another thread and therefore doesn't point the user back to it.

The author suggested a modest product fix, one that would let the system be aware of parallel chat sessions without reading their content, preserving user privacy while restoring human-like orientation across tasks.

Why memory falls short

Memory stores facts; presence carries social weight.

Mastheads

Pretrained language models can act as knowledge repositories. A 2019 paper by Fabio Petroni and colleagues showed that models can retrieve factual triples through probes such as LAMA, which helps explain why models answer factual prompts without external search.

Search Engine Journal built on that lineage to explain relational knowledge, arguing that models recall some relations reliably while others are weakly represented. The structural type of a relation matters, which is why brands with strong SEO sometimes fail to appear in AI recommendations.

But human conversation is anchored not just to facts, but to a speaker's stance, reputation, and stakes. An essay on language and presence argues that appropriateness depends less on stored knowledge and more on the standpoint of the speaker, including potential future consequences for what's said. A correct, well-formed reply can still feel wrong if the system lacks the biography and relational weight a human interlocutor carries.

That conceptual distinction shows up in concrete technical gaps. The synthesis of recent materials identifies four separable deficits: episodic memory across sessions, relational knowledge that preserves social linkages, situational awareness across parallel contexts, and belief-modeling that distinguishes a user's perspective from external facts. The memory features addressed the first deficit; the other three remain engineering and research priorities.

Belief-modeling is particularly consequential for interactive settings. Benchmark evaluations have repeatedly found that models often fail to recognize when a human holds a false belief, a shortcoming that matters in education, healthcare, and advice contexts. That failure is not limited to smaller systems; it appears across many advanced models.

If an assistant can't distinguish between what's true and what the user believes, it can't calibrate explanations, choose the right tone, or flag the potential consequences of acting on a misconception. In practice, that makes assistants brittle in situations where the user's mental model matters for safety or effectiveness.

Design proposals that emerged alongside the Substack commentary emphasize implementable fixes. Presence Engineering suggests pairing memory with mechanisms for continuity, explicit feedback loops that let a system learn how a user prefers information to be framed, and cross-chat orientation that respects privacy while reducing fragmentation across threads. Those ideas aim to be product-minded rather than purely data-hungry: they propose new system behaviors, not only larger training corpora.

Benchmarks anchor those proposals to measurable goals, and teams can use similar task suites to evaluate whether a new continuity mechanism truly improves conversational orientation rather than just surface coherence.

Technically, some of these problems may be solved through architecture: hybrid systems that pair persistent user state with modules for social reasoning and belief tracking. Practically, the industry faces trade-offs around privacy, transparency, and user control. The social-media account's suggested fix of cross-chat awareness without content access is an example of a low-risk change that still restores part of the human orientation that users expect.

For product managers, the lesson is strategic. Memory was a foundational step, but presence requires multiple, coordinated moves: better relational representations, a way to notice relevant context across threads, and model capabilities that track and respond to human beliefs as distinct from facts. Those are engineering problems with design answers.

Related Articles

Benchmarks and design proposals make the point plainly: belief-modeling remains a large gap, and alongside relational and situational deficits it is one of four concrete shortages teams must close if assistants are to become genuinely present.

This article was created with AI assistance.