Curated reading path
Understanding context
Trace how information becomes model input, survives a session, and returns after compaction or a handoff.
Follow one requirement through the system
For readers puzzled by an agent that seems to forget. Use “preserve existing term URLs” as a requirement and track where it is stored, how it is loaded, and whether it appears in the next request.
The aim is to explain a missing fact without treating every failure as a full context window. Read the ordered entries, then use the companion article’s local context-assembly experiment.
Reading order
- Token
Learn the unit used to account for text input and output.
- Context window
Separate the request’s capacity from what is actually supplied.
- Context
Identify which instructions and evidence the model can use now.
- Session
See how the application connects multiple turns without changing model weights.
- Memory system
Distinguish saved information from information retrieved for the current run.
- Context pointer
Understand why a path or URL is not the contents it refers to.
- Compaction
Recognize a compressed representation and the need to reload exact details.
- Handoff artifact
Make the next session’s starting state explicit and inspectable.
Check your understanding
A note exists in the repository but the agent contradicts it. What evidence would distinguish a failed write, a missing read, stale content, and a failure to use supplied evidence? Point to an observable file or tool result for each answer. A larger context limit alone addresses none of those four questions.