The harness matters a lot, but it's not the whole story. You're right that the memory layer isn't purely dependent on the LLM — but the LLM's ability to use a compressed summary effectively is highly model-dependent. Here's what I think is actually happening: Compaction quality is model-dependent. When Claude Code or Codex auto-compacts, it asks the model itself to summarize the conversation so far. A weaker model produces a lossy summary — it drops implicit context, forgets constraints established early in the session, and loses the thread of why certain decisions were made. A stronger model produces a denser, more faithful summary. After 10 compactions, those errors compound. GPT 5.4 and Opus 4.6 are better at lossless summarization under compression. Instruction following degrades differently across models. Long agentic tasks require the model to stay bound to the original spec even when it's no longer in the active context window. Frontier closed models appear to have been trained specifically on long-horizon task completion — they re-anchor to the original goal more reliably after context resets. Open models trained primarily on short-context benchmarks don't generalize as well to this pattern. The harness can't fully compensate. You asked whether GLM 5 in Claude Code would perform like GPT 5.4 in Codex. The answer is probably not — because the compaction summary is generated by whatever model is loaded, and the model's ability to follow that summary is also model-dependent. The harness sets the ceiling; the model determines how close you get to it. What actually helps on the open model side: aggressive context management before compaction is needed — keeping the active context lean enough that you never lose critical information in a lossy summary. I'm actually working on this problem directly in a local agentic coding tool I'm building. Two-tier approach: threshold-based compaction fires early at around 85% context targeting a watermark, so you're never compacting a bloated context. When even that reaches its limits, you go further — summarize the summaries, combine that with the original prompt and the next task list, then lobotomize the entire context and start fresh from that reconstructed state. You're not resuming the session — you're rewriting what the model believes happened. Effectively re-conditioning it from a distilled narrative rather than trying to preserve history that's already degraded. The model that handles that re-conditioning faithfully is the one that survives 10 compactions intact.