Technical breakdown of the background reasoning phase for Claude Managed Agents. Learn about self-correction and long-term synthetic memory patterns now.
What “Dreaming” Adds to Managed Agents
Claude Managed Agents already handle multi-step work with tools, state, and guardrails. Dreaming is the quieter half of that loop: a background reasoning phase that runs when the agent is not mid-turn with the user. Instead of waiting for the next prompt to notice drift, the system re-examines recent actions, open goals, and intermediate results, then consolidates what still matters into a form the agent can reuse later.
That consolidation is synthetic memory. It is not a raw transcript dump and not a permanent personal diary of the user. It is a constructed, structured record of decisions, constraints, failures, and partial solutions—written so a future step can load context without replaying every token of history.
Background Reasoning as a Separate Phase
Foreground turns optimize for latency and the user’s immediate ask. The dreaming phase optimizes for coherence over longer horizons. Practically, that means the agent can rank which threads still need attention, mark which assumptions proved wrong, and propose corrected plans before the next user message arrives. Self-correction fits naturally here: the model is free to argue with its earlier steps when the cost of a wrong path is still low.
Treat the background phase as an explicit pipeline stage, not magic. You want clear inputs (recent tool outcomes, open tasks, policy constraints), clear outputs (updated goals, memory writes, optional follow-up actions), and a hard boundary so dreaming cannot invent user intent or override safety rules that apply at runtime.
Long-Term Synthetic Memory Patterns
Useful synthetic memory is selective and typed. Store durable facts the agent verified, procedural notes about what worked in this environment, and negative knowledge—approaches that failed and why. Prefer short, labeled entries over narrative walls of text. Each entry should answer: what was tried, what result was observed, and what should change next time.
- Episode summaries: compressed arcs of a multi-step task, not full chat logs.
- Working hypotheses: current beliefs the agent is still testing, marked as provisional.
- Stable preferences and constraints: rules the user or system already confirmed.
- Error fingerprints: recurring failure modes and the fix that actually stuck.
For retrieval, rank by relevance to the active goal, not by recency alone. Stale but correct constraints often matter more than the latest partial draft. When two memories conflict, prefer the one tied to a verified tool result over pure model inference, and leave a short note that the conflict was resolved so the next run does not re-litigate it blindly.
Self-Correction Without Endless Loops
Self-correction only helps if it is bounded. Give the background phase a fixed budget of rewrites or re-plans, require each correction to cite what evidence forced the change, and stop when the plan stabilizes or when human input is required. Dreaming should reduce thrash in the next foreground turn, not open a second agent that competes with the user for control of the session.
When you design or evaluate systems like this, judge them on whether later turns start colder and finish cleaner: fewer repeated tool mistakes, clearer goal carry-over, and memory that stays small enough to load quickly. Synthetic memory earns its keep only when it makes the next action more accurate without turning the agent into a black box of private history.