The /compress 📝slash command in 📝Hermes Agent manually compresses conversation context by summarizing it, optionally keeping the most recent exchanges verbatim or narrowing the summary to a focus topic.
According to the Hermes slash-command reference, /compress here [N] summarizes everything except the most recent N exchanges, which stay verbatim (two by default), while a focus topic narrows what a full summary preserves. The CLI entry adds that the command flushes memories before summarizing. The command registry, hermes_cli/commands.py, lists compact as an alias and a --preview or --dry-run option that shows what would happen, and groups the command in /help under Context alongside /context and /status.
Context management runs on a ContextEngine, which performs compaction. The built-in ContextCompressor is the default, and a 📝pluggable context engine can replace it. The engine interface passes the focus topic from a manual /compress <focus> to compress(); engines that support guided compression should prioritise preserving related information, while others may ignore it.
The developer guide notes that a manual /compress runs a real attempt that clears the compressor's failure cooldown, and that it follows the session's compression.in_place setting. In the default in-place mode, pre-compaction turns are soft-archived under the same session id and remain searchable through session_search. Separately, the agent's automatic compressor fires at 50% of the model's 📝context window by default.
/compress shipped in release v2026.3.12 (March 12, 2026). Release v2026.3.28 preserved the transcript on /compress, and v2026.4.13 (April 13, 2026) added guided compression with a focus topic.