The Cold Start Problem

You spend forty minutes in a conversation with an AI. You've explained the project, the constraints, what failed, what you're actually trying to accomplish. The AI becomes useful — not because it got smarter, but because it finally knew enough about your situation to give relevant answers. Then you close the tab.

Two days later, you need one more follow-up. You open a new chat. It knows nothing.

This is the LLM cold start problem, and it's not an accident of bad product design. It's a consequence of how these systems work at a fundamental level.

The architecture is stateless by design

Language models don't retain information between calls. When an API call ends, nothing persists. The chat history you see in ChatGPT or Claude is maintained by the application layer — not the model. Every time you send a message, the application bundles your previous messages back into the context window and sends the whole thing again. The model doesn't remember. The application just re-explains it.

This works surprisingly well for short conversations. It starts to break when the conversation touches anything personal — your preferences, your ongoing projects, the specific vocabulary of your work, the decisions you made last month. That kind of context doesn't fit in a single session, and nothing carries it forward.

OpenAI introduced persistent memory for ChatGPT Pro users in early 2024, letting the model retain things like your name, tone preferences, and standing instructions across sessions. It's useful. It's also a compressed, opaque blob you can't inspect or query. You can ask ChatGPT what it remembers about you and get a rough answer. You can't ask it what it's inferred about your working style from six months of conversations, because the memory isn't that structured — it's more like a sticky note than a record.

Researchers are working on going further. A paper published in December 2024 described a system called Memoria that builds a weighted knowledge graph, mapping user traits, preferences, and behavioral patterns as structured entities updated incrementally over time. Genuinely interesting work. But it still treats memory as something that lives inside the AI system — inaccessible, uneditable from outside, invisible to the person whose information it is.

The session is the wrong unit of work

There's a deeper problem than the forgetting, and it's the mental model that produces it.

Most AI tools are designed around the session. You open a chat, you get something done, you close it. The session is the atom. This made sense when these tools were novelties — when every conversation was a standalone demo of a remarkable technology. It makes less sense now that people use them for real work.

Real work has continuity. Projects span weeks. Preferences accumulate. Context builds. The knowledge you need to be useful to someone isn't captured in a single conversation — it's the residue of many. The person who knows you well doesn't know you because they have a perfect memory of one long meeting. They know you because they've absorbed a thousand small things over time, without you having to re-explain them.

The problem isn't that the AI forgets between sessions. The problem is that "between sessions" is the wrong frame. Your context doesn't belong to the session. It belongs to you.

What changes when context is structured

There's a meaningful difference between memory stored inside an AI system and knowledge stored somewhere you own and can inspect.

When your context lives in a proper knowledge base — structured, named, queryable — the AI doesn't need to reconstruct who you are from scratch. It can look it up. Your current projects. Your preferences. The decisions you made with reasons attached. The people you're working with and what you know about them. None of this needs to be in the session, because it can be retrieved.

This changes the quality of AI responses in a way that's hard to overstate. Not because the model got better, but because it stopped guessing. An AI with access to specific structured facts about you can be useful to you specifically, rather than to the abstract average user it was trained on.

It also changes what "context" means. In a session-based world, context is whatever you remember to include. In a knowledge-base world, context is retrieved — selected by relevance, not by what you happened to mention five minutes ago.

The honest version of this

Structured personal context makes AI more useful. It doesn't make it accurate. The model still reasons from what you've told it, and if what you've told it is wrong or outdated, it will reason from that. An AI that knows you confidently is not necessarily an AI that knows you correctly.

There's also a version of this that goes badly: AI that silently builds a profile from observed behavior, updating preferences without asking, inferring things you'd dispute. Memory as surveillance rather than memory as record.

The version that doesn't go badly is one where the knowledge is yours — readable in a text editor, editable without ceremony, controlled by you. Where the AI can read it, but you wrote it. Where memory is visible, not inferred. Where closing a tab doesn't mean starting over.

The cold start problem is real. The solution isn't to make AI memory invisible and smart. It's to make personal context structured and portable — something you maintain, that the AI can use, and that doesn't disappear when you close a tab.

---

Asgeir Albretsen is the founder of Harbor.