📝Memory Consolidation is the process by which an AI agent distills raw session history into a small set of durable, deduplicated memories that it can load in future sessions.
Agents accumulate far more material than they can carry: transcripts, daily logs, tool results, and corrections. Consolidation reviews that material, keeps what will matter later, merges overlapping points, drops stale or one-off detail, and writes the result to a persistent store such as a memory file or database.
The process loosely mirrors the role sleep plays in human memory, and some systems name it accordingly. OpenClaw's optional "dreaming" pass scores short-term signals from daily logs and promotes only qualified items into 📝MEMORY.md, while 📝Claude Code offers auto memory that records build commands and debugging learnings across sessions.
Consolidation exists because memory competes with the task for the 📝context window. Uncurated memory grows until it crowds out the work, dilutes instructions, and repeats outdated facts. A consolidated store stays short, current, and high-signal, which separates memory that helps an agent from memory that slows it down.