Compaction Forecasting

The agent loop monitors token consumption trend and automatically saves memory 5 turns before the session would hit the compaction threshold. This ensures context is preserved without waiting for an abrupt compaction event.

How It Works

• After each turn, the loop estimates average tokens per turn from session history
• It calculates turnsUntilCompaction = (threshold − currentTokens) / avgTokensPerTurn
• Threshold is 70% of maxContextTokens
• If turnsUntilCompaction is between 1 and 5, an early memory save is triggered (postTurnAutoSave)
• No configuration required — it runs automatically

Algorithm

const compactionThreshold = agentConfig.model.maxContextTokens * 0.7
const avgTokensPerTurn = currentMsgTokens / Math.max(1, allMessages.length / 2)
const turnsUntilCompaction = Math.floor(
  (compactionThreshold - currentMsgTokens) / avgTokensPerTurn
)
if (turnsUntilCompaction > 0 && turnsUntilCompaction <= 5) {
  postTurnAutoSave(uid, userMessage, revisedContent)
}