In plain English
A model responds using the information currently available in its context. Tokens are the units that fit inside that context. Some applications may separately store conversation history or user preferences, but these features are not the same as reliable human memory.
The main ideas
Tokens
Tokens are chunks of text or other data used for processing and billing. One word may be one token, several tokens, or part of a token depending on the model.
Context window
The context window is the amount of material a model can consider during one request. Older or less relevant material may be shortened or omitted.
Conversation history
An application may send previous messages back to the model, making the exchange appear continuous.
Saved memory and retrieval
Some services store preferences or retrieve earlier records. Users should understand what is saved, who can access it, and how it can be deleted.
How it works
- 1
Collect the active context
The application assembles system instructions, conversation messages, files, and retrieved records.
- 2
Fit within limits
Material may be truncated, summarized, prioritized, or rejected when it exceeds the available context.
- 3
Generate from what remains
The model uses the supplied sequence to produce its next response.
- 4
Store or discard separately
The application—not the base model alone—decides what logs, histories, or memory records persist after the request.
Where you may see it
Long-document work
Large context windows allow more source material, but careful chunking and retrieval may still improve accuracy.
Personalized assistance
Saved preferences can reduce repetition, while creating additional privacy and correction obligations.
Ongoing projects
Conversation history helps maintain terminology and decisions, but important facts should also be kept in a dependable project record.
Cost and performance planning
More tokens can increase processing time and cost, so applications often summarize or retrieve only relevant material.
Important limits
- A model may overlook information even when it is technically in context.
- Saved memory can be incomplete, outdated, or misunderstood.
- Chat history may contain sensitive information.
- Longer context does not guarantee better reasoning or factual accuracy.
A practical reader checklist
- Keep critical project facts in an authoritative file.
- Review and delete saved information when appropriate.
- Repeat essential constraints in long conversations.
- Avoid assuming the service remembers something unless it confirms the active context.
Key takeaway
A model responds using the information currently available in its context. Tokens are the units that fit inside that context. Some applications may separately store conversation history or user preferences, but these features are not the same as reliable human memory.