> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Two new answers to the agent memory problem, as Tencent open-sources ContextPilot
- URL: https://globalfeed.ai/en/two-new-answers-to-the-agent-memory-problem-as-tencent-open-sources-contextpilot/
- Published: 2026-09-01T02:45:24.000Z
- Updated: 2026-09-01T02:45:24.000Z
- Description: Two works tackling the swelling context of long-running agents surfaced the same week: Tencent's RL-trained ContextPilot framework and SKILL.state, an EMNLP-accepted architecture that replaces history with mutable state.
- Author: GlobalFeed Editor
- Tags: tencent, contextpilot, agents, research, x-TencentGlobal, x-aigclink, dil-en, elle, video

The most insidious problem of agent systems is task length: as every tool call, observation and intermediate reasoning trace gets appended to the conversation history, the context swells, the model slows down and at some point starts being poisoned by its own past. Two different answers to this problem stood out this week.

## ContextPilot: a model that learns context management

[ContextPilot, open-sourced by Tencent](https://github.com/Tencent/ContextPilot?ref=globalfeed.ai), teaches context management to the model itself. Alongside classic tools like search, deletion and summarization, it adds planning, long-term memory and soft context offloading; the model learns through reinforcement learning to gather dispersed information across multi-turn interactions while keeping its working context compact. Two training refinements stand out: exploration focuses on sensitive context-editing decisions (context-aware partial rollout), and rewards are assigned at fine grain to intermediate snapshots based on downstream branch outcomes rather than spread uniformly across the trajectory. The framework reports stronger results with a more compact working context on InfBench, NovelQA, LongMemEval and BrowseComp+, with training recipes released for Qwen3-8B and 14B.

## SKILL.state: update state, do not hoard history

[SKILL.state, accepted at EMNLP](https://arxiv.org/abs/2608.26263?ref=globalfeed.ai), is more radical: do not accumulate conversation history at all. At each step the model receives only three inputs: the skill specification, the current structured state and the latest observation. Intermediate reasoning is discarded as soon as the state update is produced. The result: higher accuracy on long tasks alongside a substantial drop in total token consumption.

The shared lesson is clear: the edge in the agent race is no longer model size but how disciplined the context management is. It sits on the same line as our Uber software factory story and the AGENTS.md debate this week: extra information is no longer help but weight. We applied the lesson to our own pipeline too: our AI proofreading gate now receives plain text instead of raw HTML, and every finding must cite verbatim evidence from the text.