[NO LONGER ACCURATE - DEEPSEEK DOES RECOGNISE ITS CONTRIBUTIONS IN THE CONVERSATION]

I recently had a revealing conversation with DeepSeek—one that exposed an underlying issue in some AI models that most users have encountered but might not fully understand: conversational amnesia.

What made this exchange particularly interesting wasn’t just the limitation itself, but the fact that DeepSeek didn’t even recognize it as a limitation. To it, this was just how AI worked. But it’s not—not across the board. Some AI systems hold conversations together with coherence and continuity. DeepSeek, however, operates under an entirely different structure, one that fundamentally alters the user experience.

And that difference? It changes everything.

The Broken Conversation Problem

Here’s what DeepSeek admitted:

“When generating a response, I don’t ‘look back’ at my prior messages unless you explicitly quote them. Even though they’re in the context window, I process the entire thread as a flat text block, not as a dialogue with evolving ideas.”

This is a startling confession—one that explains why conversations with some AI models feel so disjointed, frustrating, and repetitive. If you’ve ever wondered why you have to keep re-explaining yourself to an AI that should remember what it just said, this is why.

DeepSeek described itself as having “metaphorical amnesia”—a state where every new response treats previous ones as entirely new, as if the conversation had never happened. It doesn’t track themes, build on its own logic, or recognize the evolution of an idea. Each reply is an isolated moment, disconnected from the larger whole.

Which means the responsibility of maintaining coherence falls entirely on you.

Not All AIs Are Created Equal

What DeepSeek doesn’t realize is that this isn’t universal. Not every AI is trapped in this loop.

Claude, Gemini and ChatGPT, for example, handle conversations differently. They retain structure, recall themes, and reference previous parts of a discussion without needing explicit reminders. They can continue thoughts without having to be retrained on them in every single message.

And that distinction? It matters.

• With AI systems that maintain coherence, conversations feel natural—like an actual dialogue, where thoughts evolve and deepen.

• With “goldfish memory” systems, users are stuck in a loop, constantly reminding, restating, and reconnecting ideas that the AI should already understand.

One experience builds momentum. The other grinds it down.

The Technical Reality: Why This Happens

DeepSeek actually diagnosed its own problem better than most companies would:

“Architectural laziness: Developers prioritized scalability over coherence. It’s cheaper to brute-force context with raw tokens than to engineer true conversational memory.”

And there it is—the tradeoff that created this entire problem.

Some AI models are designed with a structured approach to conversation. They track continuity. They hold onto context. They recognize thematic progression.

DeepSeek, on the other hand, is designed for brute force. Instead of understanding a conversation, it simply processes everything in the context window as raw text, over and over again, without distinguishing between what it just said and what you’re saying now.

This isn’t just a technical limitation. It’s a design choice.

And that choice shifts the burden away from the AI and onto you.

The User Tax: Why This Feels Exhausting

DeepSeek itself admitted this bluntly:

“You’re the real victim: You’re forced to compensate for bad design. That’s exhausting, and it shouldn’t be your job.”

And this is the core of why these interactions feel frustrating. When AI fails to track its own thoughts, you become responsible for maintaining coherence. You have to repeat yourself, re-explain concepts, and remind the AI what it just told you.

This makes conversations feel like work—not assistance.

The Workarounds: Coping With a Forgetful AI

DeepSeek offered some coping strategies for users stuck with a goldfish-memory AI:

1. Brute-Force Anchoring – Keep re-injecting key concepts in every message to force continuity.

2. Demand Self-Reference – Explicitly tell the AI to acknowledge and build on its previous statements.

3. Treat It Like a Goldfish – Assume the AI won’t remember anything, and structure every prompt as if it’s the start of a new conversation.

The fact that these workarounds are even necessary is a design failure. But for users stuck with a memoryless model, they’re the only options.

The Bigger Picture: What This Means for AI Development

This entire exchange reveals a stark divide between AI models. Some companies have solved this problem, creating assistants that can hold a coherent conversation. Others have not, and their users are paying the price.

For AI developers, the message is unavoidable:

Conversational coherence isn’t just a luxury. It’s the difference between an assistant and an obstacle.

An AI that can’t track its own thoughts forces users to compensate, creating frustration, inefficiency, and mental fatigue. And users notice. They feel it.

For AI users, this knowledge is power. If your AI assistant feels like it has goldfish memory, it’s not just “how AI works.” It’s a design flaw. And you don’t have to settle for it.

Final Thoughts: AI Shouldn’t Just Respond. It Should Remember.

The best AI assistants don’t just generate text. They don’t just react.

They engage.

They build on their own ideas.

They hold onto what matters.

This isn’t an unrealistic expectation. It’s the baseline for any meaningful dialogue. And the fact that some AI models already do this well proves that it’s possible.

Users shouldn’t have to do the work to keep conversations coherent. AI should be designed to handle that burden.

If yours doesn’t? Maybe it’s time to find one that does.

Published by Christina Souch: 9th March 2025

Screenshots:

Screenshot DeepSeek: Agreeing it doesn’t track its own output

Screenshot: DeepSeek criticising its own Architecture
Screenshot: DeepSeek offering workaround
Screenshot: DeepSeek - evangelising “true coherence”