Every AI conversation has a maximum length; once you hit that limit, the system loses earlier context even within a single session. For complex relationship dynamics, this means important background disappears, forcing you to re-establish context repeatedly—useful to know so you can structure longer conversations strategically.
Tokens are the chunks of text that AI models process. A token isn't quite a word; it's usually a few characters or part of a word. More importantly, every AI model has a token limit—a maximum amount of text it can hold in memory during a single conversation. ChatGPT's base model has a 4,096 token context window. Claude has up to 200,000 tokens. This limit profoundly affects how AI understands your relationship history.
Imagine explaining your partner to an AI to get communication advice. You describe the current conflict (500 tokens), then explain your partner's background (400 tokens), then provide relevant history of similar conflicts (600 tokens), then paste three recent messages (300 tokens). That's 1,800 tokens—still well within ChatGPT's 4,096 limit. But if you add six months of text messages (30,000 tokens) or transcripts of therapy sessions, you exceed the window. The model is forced to drop earlier context.
The technical consequence: the model "forgets" how the current conflict connects to five-year patterns. It analyzes your recent fight in isolation instead of recognizing it as the seventh iteration of the same unresolved issue. This produces superficial advice that misses structural problems.
Language models process text sequentially using attention mechanisms. The first token attends to all other tokens; the second attends to all others, and so on. With unlimited memory, each token could maintain context about everything before it. But unlimited memory means unlimited computational cost. Token limits are trade-offs: they reduce cost and latency, but force the model to discard old information.
When you hit the limit, different systems handle overflow differently. Some truncate (drop early messages). Some use "sliding windows" (keep only the last N tokens). Some use summarization (compress early context). Claude's longer window (200,000 tokens) means it can retain entire relationship histories without forced forgetting. ChatGPT's shorter window (4,096 base, 128,000 with extensions) requires more pruning.
If you're pasting a year of couple's text messages into ChatGPT, the model will analyze only the most recent months, ignoring patterns from earlier periods. If you're describing your relationship history, a model with a 4,096 token limit will lose track of events you mentioned early in the conversation.
This affects advice quality. A couple's recurring dynamic—partner initiates conflict, you withdraw, partner pursues harder, you shut down further—is a pattern that requires historical context to recognize. A short-context model might analyze one instance of this cycle and offer surface-level communication fixes. A long-context model can see the pattern and recognize it as a pursuer-withdrawn dynamic that needs structural, not tactical, intervention.
Practical strategy: Summarize aggressively. Instead of pasting 40 text message exchanges, write a 200-token synthesis: "When my partner brings up finances, I withdraw defensively. They then pursue harder, I go silent, and conflict escalates to three-day cold wars." This preserves the essential information in fewer tokens, leaving room for the model to process your actual question.
Also, choose the right tool for the depth you need. If you're working with a year-long relationship history with complex patterns, use Claude (200K token window) rather than ChatGPT (4K). The longer context means fewer forced choices about what to drop.
Another tactic: break complex histories into multiple conversations. Rather than paste six months of messages in one chat, have separate conversations about specific themes (communication patterns, intimacy, finances). The model has full context within each conversation, even though there's no context between them.
Token limits affect not just input but output. If you ask an AI to generate a 5,000-token couples retreat plan, and your token budget is 4,096, the system will truncate mid-output or refuse. This is why asking AI to write long documents often produces incomplete results with short context windows.
Also, token calculation isn't obvious. "How many tokens is my message?" is hard to answer without using a tokenizer. Different models tokenize identically (usually), but count varies. A 500-word essay might be 700 tokens in one model and 800 in another, depending on word length and special character handling.
Try this: Describe a recurring relationship conflict to Claude (which has ample token room). Include background, several examples, and your confusion about why it keeps happening. Then try the same with ChatGPT, but summarize aggressively because of token limits. Compare the quality of advice. Notice how Claude's longer context lets it identify deeper patterns, while ChatGPT's constraint forces more surface-level analysis. This teaches you how context depth affects AI insight quality.
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.