TLDR: Create a private, user-owned directory of recurring people that organizes relationship context, aliases, events, and changing facts over time, so ChatGPT can resolve who the user means and retrieve the correct current and historical context across conversations. The same architecture could also support CRM and enterprise relationship management by preserving continuity across clients, contacts, and relationship histories.
***
I’d like to propose a more structured approach to long-term memory in ChatGPT, specifically for recurring people and relationships in a user’s life.
The core idea is to create a private, user-owned “People & Relationships” layer inside memory. Conceptually, this could function like an address book or personal relationship graph. When a user repeatedly talks about the same person across different conversations, ChatGPT could associate those references with a persistent person record rather than treating each mention as an isolated memory fragment.
For example, if I frequently talk about someone I’ll call “Person A,” ChatGPT could resolve future references to that person to a specific person in my private memory context and attach relevant information to that identity: aliases or nicknames, relationship to me, communication style, important events, preferences, recurring patterns, current relationship state, and dated historical context.
The key difference from ordinary persistent memory is that this would be person-centered and temporally aware.
Instead of memory containing disconnected facts such as:
“Person A prefers X.”
“Person A and I had an argument.”
“Person A now prefers Y.”
the system could model them as part of an evolving timeline:
“Person A preferred X during this period.”
“A later event changed the relationship or preference.”
“The current known state is Y.”
Older information would not necessarily be deleted, because historical context can still be relevant, but newer information should be able to supersede older information for current-state retrieval.
This would help avoid a common problem with long-term memory systems: stale facts resurfacing after they are no longer current. Chronology should be part of the memory model, not just metadata attached to otherwise timeless facts.
From a technical perspective, I imagine something like a private entity graph with temporal state.
Each recurring person could have:
- A stable user-local entity ID
- Name and aliases
- Relationship type or description
- Source attribution for memories
- Timestamped observations and events
- Current-state attributes
- Historical attributes
- Confidence or uncertainty levels
- Supersession links between older and newer facts
- Relationship-specific communication patterns
- Links to relevant conversations or events
Retrieval could then work in two stages:
- Entity resolution: determine which person the user is referring to.
- Context retrieval: retrieve the most relevant current and historical information about that person for the present conversation.
This would be especially useful in new conversations. If I mention “Person A,” a nickname, or another recurring person, ChatGPT could identify who I mean and retrieve the appropriate relational context without requiring me to restate years of background.
Importantly, this should remain entirely user-side. I am not suggesting that ChatGPT globally identify people across different users or combine information from separate accounts.
My representation of Person A should belong only to my personal context. If Person A also uses ChatGPT, their private conversations should not be available to me, even indirectly.
The system should model “what the user has shared and observed about this person,” not claim to possess an objective global profile of that person.
This distinction is important for privacy. A global identity system could create serious problems where information shared privately by one user influences responses given to another user. A user-local relationship graph avoids most of that issue while still providing richer continuity.
The feature should also distinguish between different epistemic types of information.
For example:
- Explicit user-provided fact: “Person A told me they dislike X.”
- User observation: “Person A tends to respond this way.”
- Model inference: “Based on several interactions, Person A may prefer X.”
Those should not all be stored or presented with the same confidence. Model inferences should remain clearly distinguishable from facts supplied by the user.
There should also be a clear difference between:
- Replace this fact
- This fact changed over time
- Keep this only as historical context
- Forget this entirely
These are different memory operations.
The broader goal is to make long-term relational context feel more natural.
Humans do not understand long-term relationships as flat collections of facts. We understand people as evolving individuals with history, changing preferences, repeated patterns, and current states.
Someone I have known for many years is not represented in my mind as a list of disconnected memories. My understanding of that person is continuously updated and compressed into a relational model.
ChatGPT does not need to store every conversation verbatim to approximate that. A structured, temporally aware person model could provide much of the same continuity while remaining computationally and privacy-conscious.
This could significantly improve conversations about:
- Friends
- Partners
- Family members
- Coworkers
- Managers
- Clients
- Doctors
- Recurring professional contacts
- Long-running interpersonal situations
Business / CRM use case:
The same architecture could also support business relationships such as clients, prospects, vendors, managers, recruiters, or other recurring contacts. A person-centered temporal memory model could preserve communication preferences, role changes, prior commitments, relationship history, and current state, giving employees immediate relational continuity when working with someone. In practice, this could create a form of “instant rapport” by letting ChatGPT or a CRM user begin from an organized history rather than repeatedly reconstructing the relationship from scattered notes.
It could also improve pronoun resolution, nickname resolution, name ambiguity, and retrieval across separate chats.
From a UX perspective, this could simply appear as a “People” or “Relationships” section in memory.
Users could inspect and manage each person profile, including:
- Merge duplicate people
- Separate incorrectly merged identities
- Correct names or relationships
- Review current versus historical information
- Remove individual facts
- Delete the entire person profile
- Disable automatic person memory
The interface could remain simple even if the underlying architecture is more sophisticated.
I think this would make long-term conversations substantially richer than simply increasing the number of saved memories.
The important distinction is:
More memory is not the same thing as better-organized memory.
A person-centered, temporally aware, user-owned relationship memory system could allow ChatGPT to maintain continuity across years while avoiding the privacy problems of cross-user identity matching.
In technical terms, I think the underlying concept is closest to a user-local temporal knowledge graph or personal relationship graph with entity resolution, provenance, confidence, and state transitions.
The user-facing version, however, could simply feel like an intelligent address book that helps ChatGPT remember who the important people in your life are and how those relationships have evolved over time.