Retrieval lets the agent reread the past. Consolidation takes those experiences and turns them into something the agent can act on without rereading every episode. One scales the lookup. The other is what makes the agent better at its job.
Anthropic shipped Dreams this month as one approach. A dream takes a memory store and past session transcripts, produces a candidate consolidated state, and gives the user the veto before any of it becomes default. Making the consolidated state a separate reviewable artifact with a human reviewer is one design solution, not the default way to do it.
Three questions are worth discussing before any agent runs consolidation in production.
- Does consolidation happen at all, or is the system just rewriting the filing cabinet?
- Are the inputs safe, or can one injected episode become a durable lesson?
- Does the consolidated output match what the user actually taught the agent, or does the agent's own loop quietly overwrite what the user tuned?
I do not have a clean answer to the third one yet. I wrote two longer pieces on my site that walk through where I have landed so far.
It has also occurred to me that if you have workflows or skills that learn from previous actions, problems, or resolutions, you are already doing some level of consolidation. It just may not be automatic.
What guardrails would you want before you let your agent rewrite its own memory?
Written by Duane Grey
AI Strategy & Implementation
Independent AI consultant helping companies cut through hype and deploy systems that produce real results.