Recent messages are kept in high detail close to raw, older messages get semantically summarized rather than dropped outright, and stable facts or preferences may be stored as long-term memory retrieved across sessions. The final context sent to the model is assembled from these pieces, not the full conversation word-for-word.

This is not exact compression like ZIP, where the original text can be perfectly restored — it's closer to semantic summarization. The system tries to preserve important meaning, decisions, user preferences, project context, and repeated facts, while dropping low-value details like typos, casual comments, and outdated intermediate decisions. For example, a long conversation about building a site called Nextforay on GitHub Pages versus Cloudflare Pages, with a private repo and specific branding preferences, might compress down to a two-sentence summary that keeps the decisions and preferences but drops the back-and-forth that led to them. That means it's lossy: exact wording, code, YAML, DNS records, or precise numbers can get lost or simplified during summarization. For technical or high-precision tasks, it's safer to re-paste the latest exact version than to rely on the model's memory of it.