openclaw-mem

To preserve memory across session compaction, important knowledge is separated into durable disk files like MEMORY.md and daily logs, while pre-compaction flushing ensures transient context is saved before the window is trimmed.

The best way to separate memory is to keep short-term session context lightweight for transient chat, while writing lasting facts into curated MEMORY.md entries and day-specific work notes into daily log files.

Retrieving decisions requires retrieval-first recall using memory_search and memory_get functions to query indexed durable memory, ensuring reliable recall of past preferences and work history.

Yes, the memory management approach includes privacy-safe handling of sensitive information, ensuring that durable knowledge and daily work notes are stored cleanly and securely across compaction events.

You need to flush session memory to disk during pre-compaction flushing, which must occur right before the context window gets trimmed to prevent losing active decisions and temporary work notes.

Agents forget preferences after context trimming if durable memory is not properly indexed and flushed, making retrieval-first recall and curated MEMORY.md entries essential for reliable recall.