Memory

Continuity has two scales. Within a run, the run's own session is the context, so a planner giving a developer another task sees its recent work. Across runs, continuity is bounded memory: a compacted Markdown summary per scope, injected into briefs and refreshed after runs, so recurring work does not replay an ever-growing session.

Scopes and storage

Memory lives under .polyglav/memory/ in the worktree, one file per name:

Scope	Path	Refreshed	Injected
role	.polyglav/memory/roles/<type>.md	After a delegated agent run	Into that role's sub-agent system prompt
team	.polyglav/memory/teams/<name>.md	After a team run	Into every stage brief
job	.polyglav/memory/jobs/<name>.md	After a job run	Into the run's system prompt

Files are plain Markdown and human-editable. Reads fall back to the older locations (.polyglav/jobs/<name>.memory.md, .polyglav/teams/<name>/memory.md) so existing memories keep working, and writes always go to the new path. Writes are atomic (a .tmp file replaces the target).

Automatic and manual

Each scope is refreshed by summarizing the run through the same compaction path as /compact, seeded with the previous memory so facts carry. A failed summarizer leaves the previous memory in place.

• memory (default true) enables or disables all scopes.
• memory_scopes (default all true) toggles a scope on or off individually.
• memory_max_chars (default 2000) caps the summary text injected into a prompt or brief.
• /memorize summarizes the active run into the active role's memory, and /memorize <role|team|job> <name> writes a named scope.

See config.md for the keys and teams.md / jobs.md for the pipeline and scheduler behavior.