Anthropic launched Dreams API (research preview) and Outcomes API (public beta) as part of Claude Managed Agents, announced at Code with Claude 2026 (May 6).
Dreams API is a formal inter-session memory consolidation primitive. It reads an existing memory store + up to 100 prior sessions and writes a new output memory store (input store is not modified). It reorganizes memories, merges duplicates, replaces stale entries, and surfaces patterns. Runs on claude-opus-4-7. The recommended architecture is a three-store model: (1) read-only org store, (2) read-only project store, (3) read-write working store with human-review promotion gates. This is structurally identical to TITAN's hot/warm/cold tiering with /dream promotion.
Outcomes API implements a rubric-gated self-correction loop. API surface: client.beta.sessions.events.send() with event type "user.define_outcome" including description, rubric content, and max_iterations (example: 5). A separate grader LLM evaluates the artifact against the rubric and returns gaps to the working agent, repeating until criteria are satisfied, budget exhausted, or outcome fails. Anthropic reports up to 10-point task success improvement vs standard prompting. No surcharge — standard $0.08/session-hour Managed Agents rate applies.
Security risk: Memory stores are persistent influence channels. Prompt injection into a working store propagates to all future sessions that read it. Mitigate with: immutable org/project stores + explicit promotion workflow with provenance checks before writing to org store.