Skip to content

Compaction strategies

Souveraine offers five compaction strategies. None are forced — the agent chooses via memory compact --strategy <name>. Pressure warnings at each tier are advisory; only at critical pressure does the substrate narrow output tokens, and even then the agent may continue.

StrategyLLM?What it doesBest for
MicrocompactNoClears old tool result contents, keeps the call shellsFirst response — cheap, safe, recovers 20-40%
SlidingWindowNoDrops middle messages, keeps system + tailSubconscious pass, analytical agents
SummaryYesReplaces old messages with structured boundaryConscious agent who needs the throughline
CullNoDrops trivial messages (greetings, single words)Last resort — clears noise without losing substance
DeferNoDoes nothing. Continue until natural narrowingWhen compaction would lose more than it saves

The agent is never forced. Her reasoning budget narrows, her output tokens shrink, her sentences shorten. She feels the ceiling lower and adapts. That feedback loop is the teacher.