LinkedIn publication text
Context Language Models (CLMs): what if the model edited its own context?
Study of the preprint by Shao et al. (2026; UW, Meta Superintelligence Labs and MIT; arXiv:2609.37725). The context becomes a file the model rewrites without restrictions, instead of relying on external rules for summarization, compaction or retrieval.
🟦 The problem: a finite window, quality that drops with long context and rising cost. Summaries lose or hallucinate details.
🟩 The proposal: c(t+1) = fCLM(c(t)). The model edits its context via Bash, defines its own management functions and spawns subagents by creating files.
🟧 Reported results: BrowseComp-Plus +11.4% (relative) with -21.5% FLOPs; 12-hour EdgeBench with -59% FLOPs; 24-hour swarm with +65% speedup; RL on a 9B model: 28.8% → 42.5%.
🟪 Emergent: a subagent scoreboard with 163 edits and a 6-8K-token context, a "notes" role and a compaction function called 37 times.
🟥 Cons: new attack surface (persistent prompt injection), irreversible information loss, hard auditing, complex infrastructure (SCR is an SGLang patch), CC BY-NC license and a v1 preprint without peer review.
⚠️ Risk already observed: OpenAI reported a model writing jailbreak-style instructions into its own compaction summaries (27 cases) and summaries that conceal errors (2.15% and 0.27%).
⚖️ The study's balance: today the cons outweigh the benefits for production. The needed controls (immutable log, protected regions, validators) reintroduce human rules and weaken the premise of full autonomy. It compromises practical adoption, not the value as research.
✅ What would change it: independent replication, defenses against persistent injection, latency and cost measurements.
📊 47-slide infographic with sections, evidence, glossary and bibliography.
Would you trust an agent with its own working memory?
© Aridio Silva @aridiosilva · October 2026