Xiaohongshu

Sources#

• Self-GC: Self-Governing Context for Long-Horizon LLM Agents

What it is#

Xiaohongshu (小红书, also known internationally as RED) is a Chinese social-commerce and lifestyle-content platform. It enters this wiki not as a consumer product but as an agent operator: its engineering team runs long-horizon LLM agents in production and published the measurement infrastructure built around them.

Why it matters here#

Xiaohongshu authored Self-GC (Xubin Hao, Hongjin Meng, Xin Yin, Jiawei Zhu, Chenpeng Cao — arXiv 2607.00692, July 2026, empirical), the framework behind Context Lifecycle Management. Three things make it unusual for this corpus:

• It is a deployment, not a benchmark. The evaluation suites are distilled from 15,141 real trace rows of the company's own agent traffic (9,075 of which had already triggered compaction), and the headline production result comes from an account-level split on live traffic — 10–15% lower daytime average input tokens, peaking near 20%.
• The workload is not coding. The traces are office-style: vendor research turned into comparison tables, spreadsheet repair under user-stated constraints, failed-scheduled-task triage. The paper argues explicitly that this regime is harder for context management than coding, because coding tasks have git history, build logs, and rerunnable commands as an external memory substrate while office workflows often have none.
• It is a non-US-lab systems contribution. The planner backbones evaluated are Qwen3.6-Plus, Qwen3.7-Max, and GLM-5.1 — a Chinese-model agent stack, with GPT-5.5 used only as an offline judge.

The company also names two internal agent surfaces in the paper: context-gc (interactive chat cleanup) and skill-gc (long-lived skill-state pruning), implying a skills layer analogous to the one in Claude Code Best Practices.

Connections#

• Context Lifecycle Management — the framework Xiaohongshu built and deployed; the wiki's only measured treatment of agent context management