Fold

1. Start with raw messages

A continuous stretch of context can be written as m1 through m10. These are simply the original messages, in order.

As the task continues, raw messages keep accumulating: m1, m2, m3 … m100 … m1000. The model does not need every historical detail to stay in its current context.

2. Digest

Suppose m1–m4 describe one completed stretch of work. We can compress the important information into a shorter account. This compressed content is called a digest.

A digest keeps the main information needed to continue working, while leaving out details that no longer need to stay in context.

3. Chunked compression

We do not compress all of m1–m10 into one digest. Instead, divide the history into local stretches and compress each completed stretch separately.

This keeps the work from different stages independent. The two completed stretches are not merged into one global digest.

4. Source Reference

A digest alone is not enough. Suppose d1 says “Fixed the routing bug. Tests passed.” Later, the model asks: “What exact timeout value was used?” That detail may be missing from the digest.

Each compressed block therefore also records which original messages it came from. This record is its source reference.

When an exact detail is needed, follow the reference to read the original content. The original messages remain available; the digest does not have to contain every detail.

Fold

Now we can define the complete result. A folded block contains both the compressed content and a reference to its source.

Fold reduces active context and keeps a path back to the original context.

message → digest → chunked compression → source reference → Fold block