Long conversations

When a conversation approaches the model’s context limit, FreeCode folds older turns into a structured summary and keeps the most recent turns verbatim. It fires slightly before the window is full so there is always headroom for the reply, and a PreCompact hook can defer it.

What comes back after a compaction

The summary replaces the full text of everything older, including files the agent read. So the agent does not have to re-read what it was working on, FreeCode attaches three things to the summary, the way Claude Code and pi do:

• The 5 most recently read or edited files, as they are on disk now (up to about 5K tokens each, 50K in total). A file whose contents are still in the kept recent turns is not repeated.
• The paths of every other file touched in the session, split into modified and read, so the agent knows where to look.
• Skills it loaded earlier (up to about 5K tokens each, 25K in total), since their instructions still apply.

After a compaction, reading a file is always real again: the “unchanged since you last read it” shortcut only applies to a copy the agent can still see.

Set FREECODE_COMPACT_RESTORE=0 to turn the attachments off.

On this page (coming soon)

• When compaction triggers (limit minus a safety buffer)
• What is preserved verbatim and what gets summarized
• The summary’s structure: Goal · Done · In Progress · Blocked · Decisions · Files · Next Steps
• Forcing it early with /compact
• The fallback when the summarizing call fails
• Tuning the knobs — see compaction internals