Sessions and context compaction
Preserve channel-scoped conversations while keeping long-running context within model limits.
Session keys
Keys scope a conversation by agent and channel, with optional group or peer identifiers. Channel isolation prevents unrelated conversations from sharing transcript state, and sub-agents receive separate scoped sessions.
Database-backed transcripts
Messages append transactionally to durable session records. Cursor-based pagination supports long histories, while database persistence allows the daemon to recover continuity after a restart.
Automatic compaction
Compaction triggers when total context exceeds the model context window minus reserved response tokens. The pipeline preserves priority context, summarizes older exchanges, flushes durable memories, and records outcome metrics.
Soul synchronization
Soul sections are assigned priorities so core truths and boundaries survive aggressive trimming. File watching and BLAKE3 change detection synchronize identity updates without treating the Soul file as the transcript store.