What's this about?
Safety from within, through memory-native state evolution — safety as a property of how the model's persistent state evolves, rather than as a filter applied at the boundary. 20 upvotes.
What is "memory-native" safety?
Current safety is mostly perimeter defence: check the input, check the output, refuse when something matches. That works on a single turn and degrades over a long session, which is exactly what the Compaction Cliff paper showed — a rule that must be followed verbatim gets summarised away while the content it constrains grows. Making safety a property of the evolving state means the constraint is maintained as part of what the agent *is* across the run, not re-checked at each edge. Whether that is achievable is the open question; the framing is the contribution.
Why does this matter to you?
- It targets the failure mode this digest has now covered three times: guardrails that quietly weaken as context fills.
- "Safety from within" versus filtering is the same structural argument as authorization versus detection for prompt injection.
- Long-running agents are where this matters, and background computer use shipped today.