Why does ChatGPT keep confusing two people in my life?

Because its memory layer doesn't model entities. Names, contexts, and facts about people get stored as loose text. Two people named Sarah, or a co-founder Alex and a friend Alex, end up sharing one fuzzy profile.

Is this an LLM problem or a memory problem?

Memory. The base model is actually decent at entity resolution within a single conversation. The problem is what gets stored: a flat string like 'works with Alex on the product launch' doesn't disambiguate which Alex, and the next session has no way to tell.

Does giving people unique nicknames in chat help?

Yes, materially. If you consistently say 'Alex (co-founder)' and 'Alex (cousin),' the model is much more likely to keep them separate. It's a workaround, not a fix, but it works.

Why does this matter beyond annoyance?

Because the assistant is increasingly used for drafting messages, summarizing conversations, and giving relationship advice. Fusing two people produces drafts addressed to the wrong person and advice grounded in a portrait of someone who doesn't exist.

What's a per-entity atom?

Each person in your life becomes a distinct entity with a stable ID. Facts about them are atoms attached to that entity. New mentions of 'Alex' get resolved to a specific Alex (or flagged as ambiguous) before any fact gets stored. The model can then say 'which Alex?' instead of guessing.

Does Konshus do this?

Yes. Entity resolution is built into the distillation loop — atoms about people are attached to entity records, and ambiguous mentions are surfaced for you to resolve before they pollute the persona. Per-entity correction is covered by the patents filed in 2026 (/patent).

What about ex-partners or estranged people I don't want surfaced?

Per-entity controls let you mark an entity as private or archived, so the assistant won't surface them in suggestions or drafts unless you explicitly ask. That's the upside of having entities at all.

A guide · ~8 min read

Why Your AI Confuses Your Partner, Co-Founder, and Ex

The assistant drafts a message to your co-founder Alex and accidentally references something your cousin Alex said at Thanksgiving. It summarizes a call with your partner and mentions a project your ex was on. It's not making things up — it's just unable to keep the people in your life separate. That's an entity-resolution failure, and it's structural.

What entity resolution actually is

Entity resolution is the process of figuring out whether two mentions refer to the same real thing. "Alex" in chat A and "Alex" in chat B — are those the same person? The base language model can usually do this within a single conversation by reading context. The memory layer downstream doesn't, because it stores text snippets without entity IDs. So once two different Alexes are in your memory store, the system has no reliable way to keep them apart.

The three failure patterns

Name collision

Two people with the same first name get fused. Their facts pool. The assistant cheerfully attributes one's job to the other.

Role drift

The role someone played changes over time — your manager becomes your peer, your contractor becomes a full-time hire. Without temporal context, old roles linger and the model treats them as current.

Relationship erasure

Estranged people, ex-partners, former co-founders — the assistant has no notion of "this person is no longer part of my life" and will helpfully suggest you reach out. This is the most painful version of the failure. (For the broader contradictions case see why your AI contradicts itself.)

Short answer

AI assistants confuse the people in your life because their memory layer stores facts as flat text without entity IDs. The fix is per-entity atoms: every person gets a stable ID, ambiguous mentions are resolved before storage, and you can edit, archive, or hide entities individually. Per-entity correction is one of the patents filed in 2026 (/patent).

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus

What a per-entity model looks like

Each person in your life is an entity with a stable ID, a display name, optional aliases ("Alex (co-founder)"), and a relationship type (partner, family, colleague, friend, former). Atoms about them are attached to the entity, not floating in a memory blob. When a new chat mentions "Alex," the system either matches confidently or surfaces the ambiguity to you: "which Alex?"

That single design choice eliminates most of the pain. The assistant drafts messages to the right person. Summaries attribute things to the right person. Estranged or archived entities don't surface unless you explicitly ask. The model gets more useful because the memory layer finally has the structure the model can lean on. (We argue the broader per-claim version of this in why your AI says weird things.)

Workarounds that help now

Use disambiguating tags consistently. "Alex (co-founder)" and "Alex (cousin)" everywhere. It feels formal; it pays off.
Keep a relationships doc in your custom instructions. Short list of the people who matter, with one-line descriptions. Paste at the top of any chat that involves drafting or summarizing.
Clean up Memory when relationships change. When someone exits your life, scan the Memory panel and delete entries about them. Otherwise they'll surface for years.
Use a memory layer with entities. The only real fix. Konshus is built this way; a handful of other tools are getting there.

Frequently Asked Questions

An assistant that knows your people apart

Konshus models every person in your life as a distinct entity with a stable ID, so the assistant stops fusing your co-founder with your cousin. Patent pending — see /patent.

Meet Konshus