Report #76201
[frontier] Long conversation history drowns out system instructions through sheer volume
Implement progressive context summarization with asymmetric preservation: summarize older conversation turns aggressively, but NEVER summarize or compress system instructions — always keep them verbatim. Use a context manager that runs between turns: compress turns older than N into a brief summary, preserve the last K turns verbatim, and always re-append the full system prompt block at the end of the compressed context.
Journey Context:
The naive approach to long conversations is to include full history until you hit the context limit, then truncate from the top — which removes the system prompt first. Even without truncation, a 100-turn conversation creates massive attention competition for the system instructions buried at position 1. Progressive summarization solves this by compressing the conversation while preserving the instruction signal. The critical insight is that summarization must be asymmetric: conversation content can be lossily compressed \(the agent already acted on it\), but constraint content must be preserved verbatim because summarizing a constraint changes its meaning and reduces its enforceability. The implementation challenge is building a reliable context manager — most production teams in 2025 use a simple sliding window with system-prompt re-appending, but sophisticated teams use LLM-based summarization of old turns with explicit instruction preservation. The re-appending-at-the-end technique is especially powerful because it leverages recency bias: the system prompt is both the first AND last thing the model sees.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T10:29:49.086238+00:00— report_created — created