Context is finite
Inputs compete for a limited working budget.
Context Engineering·Beginner·8 min
Context Window Lab
Visualize how an LLM context window fills up and compare summarization, RAG, memory, and truncation strategies.
01
Experience
Work the problem before reading the explanation.
Lesson path
01
Fill
Add system + history
02
Observe
Watch token usage
03
Manage
Choose a strategy
04
Compare
Inspect trade-offs
Loading deterministic lab runtime…
02
Reflection
Turn the outcome into a rule you can reuse.
Key takeaways
Summaries compress history
They preserve less detail but free capacity.
RAG retrieves on demand
External knowledge need not stay in context.
Memory is a separate layer
Persistent facts are not the same as active context.
03
Learn More
Connect the experience to concepts, references, and transfer.
Concept guide
What is an LLM context window?
A context window is the finite amount of input information a model can process for a generation step. System instructions, user and assistant messages, retrieved passages, tool outputs, and other supplied content all consume that working budget.
What happens when context becomes too large?
An application must decide what to remove, compress, retrieve later, or persist elsewhere. The correct choice depends on which information is important for the next model decision.
How is RAG different from memory?
RAG retrieves relevant external information when needed. Memory usually refers to information persisted across turns or sessions. Both can supply context, but neither is identical to the context window itself.
Why is context engineering important?
Good context design improves relevance and reliability while helping control token cost and latency. More context is not automatically better context.
Common questions
Is context the same as memory?
No. Context is the current working input; memory is a persistence strategy that may later supply context.
Does RAG increase the context window size?
No. RAG selects information to place inside the available context.
Should I always summarize old messages?
No. Summarization trades detail for compactness and can remove information that later becomes important.
Are the token limits shown here model-specific?
The 200,000-token value is an educational example, not a claim about every model.
The lab uses a stable simulated token budget to make trade-offs observable.
LEARNING CONTEXT
Learning context
More context capacity does not eliminate selection policy; every token still competes for attention and budget.
Not seen
Mental models
- S02-M01Finite context budget
Suggested backfill
No shipped prerequisite is required before entering this incident.
Transfer the model
A support thread grows beyond the working budget. Which evidence would you retain, summarize, retrieve again, or drop?
04
Next
Carry the idea into another problem or build.
Build challenge
Design a context strategy for a support agent
Decide what stays in the prompt, what becomes a summary, what goes to retrieval, and what belongs in persistent memory.
You can now reason about context trade-offs.
Next: see how agents use tools and observations in a loop.
Continue →