Why Your AI Forgets You

Open ChatGPT. Tell it your name, your job, what you're working on. Close the tab. Open it again tomorrow. It has no idea who you are.

This is the default state of every AI assistant: stateless. Each conversation is an island. The model might be brilliant at reasoning, but it can't remember that you told it your dog's name last Tuesday.

The Problem Isn't Intelligence

Large language models don't have persistent memory by design. They process a context window — a fixed-size buffer of text — and generate a response. When the conversation ends, that context is gone. There's no mechanism to carry information forward.

Some products paper over this with conversation history. They feed previous messages back into the context window. But context windows have limits, and this approach doesn't scale. You can't fit six months of conversations into 200K tokens.

What Real Memory Looks Like

When I built Edward, the first problem I wanted to solve was memory. Not “save the last 10 messages” memory — real, long-term, searchable memory that works the way yours does.

Every conversation is mined for memorable information. Edward identifies facts (“user's dog is named Luna”), preferences (“prefers dark mode”), context (“starting a new job Monday”), and instructions (“always respond in bullet points”). These get stored with vector embeddings in PostgreSQL using pgvector.

When you start a new conversation, Edward runs a hybrid retrieval — 70% vector similarity, 30% BM25 keyword matching — to pull in relevant memories. The result is an AI that actually knows you. Not because it memorized your chat logs, but because it extracted and indexed what matters.

Memory Types Matter

Not all information is created equal. A fact about your allergies is more durable than the context that you're “on vacation this week.” Edward classifies memories into four types — fact, preference, context, and instruction — each with different extraction patterns and retrieval weights.

This matters for long-term accuracy. An AI that treats “I'm feeling tired today” the same as “I'm allergic to shellfish” will eventually surface stale context when you need critical facts.

Beyond Retrieval

Memory retrieval is step one. Edward also runs background consolidation — an hourly process that clusters related memories, flags stale information, and creates connections between facts. Think of it as the AI equivalent of sleeping on a problem.

There's also a reflection system that generates follow-up queries after each conversation turn, enriching future context with memories you didn't explicitly ask about but that are relevant.

The Takeaway

An AI without memory is a tool. An AI with memory is an assistant. The difference isn't in the model — it's in the infrastructure around it. If you want an AI that actually knows you, memory can't be an afterthought. It has to be the foundation.