Report #38160
[frontier] Agent forgets system prompt personality in long context window
Implement Context Hydration by re-injecting the core identity and constraint block at the end of the conversation history right before the model generates its next response.
Journey Context:
Practitioners assume larger context windows solve memory, but they actually cause attention dilution. The model pays exponentially less attention to token 10 when there are 100,000 tokens. Re-injecting the identity block forces it back into the local attention window, treating the system prompt as both a prefix and a suffix constraint.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T18:31:51.771805+00:00— report_created — created