Scroll reframes agent context as code

This is a stronger abstraction than another fixed memory schema: recall becomes open-ended computation over lossless history, though production use will depend on safe sandboxing and reliable state management.

• –Preserves ground-truth history instead of committing early to lossy summaries
• –Uses typed variables to keep tool outputs and derived state out of repeated prompts
• –Evicts stale context while retaining direct links back to exact historical events
• –Reports 94.8% on LongMemEval_S and 73.1% on BEAM_10M with Qwen3.8-Max
• –Signals a shift toward programmable agent harnesses rather than static retrieval pipelines