Memory and Context

Clouisle has three layers of context: the current active message branch, conversation-level summaries, and cross-conversation user memory. Each solves a different problem.

Enable Cross-Conversation Memory

1. Open the Agent's Orchestration > Memory.
2. Turn on Enable Memory.
3. Set the maximum number of memories per retrieval, range 1-50, default 10.
4. Turn on Auto-extract Memories as needed; it is on by default.
5. Save the Agent and verify in two different conversations that user preferences are reused.

How Long Conversations Are Compressed

Context compression is on by default. It computes the available input budget from the model's context window, reserving a default of 4000 output Tokens and a 1000 safety margin.

• At 70% utilization, warning pressure begins.
• At 80% utilization, selective compression is triggered.
• At 92% utilization, it escalates to blocking-level macro compression.
• By default, the last 3 raw turns and 2 tool turns are kept.
• On context length errors, the system performs at most one more aggressive retry.

The streaming API can send compression_start and compression_end to explain why this turn reorganized the history.

Conversation Summary

A conversation-level summary is extracted asynchronously after at least 4 turns by default, with an injection budget of 400 Tokens by default. After 3 consecutive failures, the circuit breaker opens and cools down for 600 seconds. The summary only serves the current conversation and does not replace cross-conversation memory.

Troubleshooting

• Preferences not remembered: Confirm the Agent has memory enabled, the information reached the importance threshold, and the same signed-in user is used.
• Replies ignore very early details: Long conversations may have been compressed; restate the key constraints or start a new conversation.
• Compression happens too often: Check the model's context length, output reservation, and the size of tool results.