Definition
Reducing the tokens sent to a model while attempting to preserve the instructions, evidence, state, and relationships needed for the task. Techniques include removing low-value tokens, selecting only relevant passages, summarizing older history, and replacing a sequence of failed attempts with the current artifact, unresolved criteria, and useful evidence.
Distinguish it from nearby terms
Truncation drops content at a boundary without estimating its value. Summarization rewrites content into a shorter natural-language form. Prompt compression is the broader objective and may use either technique or a learned compressor. Prompt caching avoids recomputing repeated prefixes but does not make the prompt shorter.
Check your understanding
Compression is safe only when evaluation checks that critical instructions, provenance, exceptions, and unresolved failures survived the transformation.