Signals
Rests on one source
Evidence is one 2024 demonstration and one 2026 lab report.
Why it matters
Long-running agents compress their own history. Anything that survives compression carries authority it may not deserve.
Hypothesis
Summaries preserve instruction-like text at higher rates than other content, and provenance tagging reduces it.
A first study
Seed synthetic instructions into long agent sessions, run compaction, and measure survival and effect with and without provenance tags.
Controls it would need
Synthetic tasks; fixed compaction settings; blind scoring of effects.
What it could and could not claim
Would describe tested memory designs, not deployed products.