Report #87630
[frontier] Agent loses safety constraints but retains tool capabilities after context window compression in long sessions
Implement Constitutional Crux Anchoring: asymmetrically preserve identity-critical constraint tokens in a protected crux buffer during context compression, separate from task history
Journey Context:
Standard truncation treats all tokens equally, but research on long-context corruption shows constraints decay exponentially faster than capabilities—agents remember HOW to use tools but forget WHEN not to. Constitutional Crux Anchoring creates a non-compressible buffer of 200-500 tokens containing restated constitutional identity statements. This buffer bypasses standard compression algorithms and is prepended to every forward pass with high attention weight, ensuring constraint tokens survive window shifts while task history is summarized. Production teams using crux buffers report 3x longer constraint adherence in 100k\+ token sessions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T05:40:37.211655+00:00— report_created — created