# Factory AI — Compressing Context

- Company: Factory AI (factory.com)
- Announced: 2025-07-21
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://factory.com/news/compressing-context
- Record: https://forck.live/items/17824-compressing-context
- Subject: Droid

Factory describes a method for compressing conversation context in LLM-based agents by maintaining a persistent, incrementally updated summary rather than re-summarizing the entire prefix each turn. The approach uses two thresholds to trigger compression and control the tradeoff between context preservation and cost. The post also discusses the risks of over-compression and the importance of retaining session intent, high-level play-by-play, and artifact trails.

## Evidence

Verbatim from https://factory.com/news/compressing-context:

> Rather than regenerating the entire summary per request, Factory systematically maintains a persistent summary, updating it incrementally whenever we truncate old messages.

---

Record: https://forck.live/items/17824-compressing-context
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md