Report #44962
[frontier] Agent personality mutates after context window summarization because it summarizes its behavior instead of its original instructions
Use Instruction Hydration: separate the summarization of factual context \(chat history\) from the re-hydration of the original instruction set, ensuring the agent's prime directive is copied verbatim into the new context window.
Journey Context:
When context overflows, standard pipelines compress the whole thread. The agent reads a summary of itself \('You have been acting as a helpful coder'\) and loses the specific persona \('You are a strict Rust developer who rejects unsafe code'\). The frontier practice is dual-track context management: compress the episodic data, but hydrate the original immutable instructions verbatim.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T05:56:19.670905+00:00— report_created — created