Claim
A scheduled background process reads past session transcripts and memory stores, extracts cross-session patterns (recurring mistakes, emerging best practices, stale memory, contradictions), and proposes curated memory updates. Updates require human review and approval before applying.
Mechanism
Single-session memory loses cross-session signal. A process that reads across sessions surfaces patterns no individual session can see. The human review gate prevents drift: proposed updates are visible, reversible, and auditable before they change the agent's operating context. This is proposed memory, not automated memory.
Conditions
Holds when: session transcripts are stored and accessible. Cross-session patterns exist because the agent runs repeatedly on similar tasks. A human can review proposals on a cadence that matches session frequency.
Fails when: sessions are too infrequent or too varied for patterns to emerge. Review is skipped, turning the process into unreviewed automated memory.
Evidence
Announced as a gated research preview at Code with Claude, May 6, 2026. The feature reads past session transcripts and memory stores, extracts cross-session patterns, and curates memory updates with the option to review changes before they land.
Signals
- Proposed updates cluster around the same mistake class week-over-week (root cause, not symptom)
- Human reviewers approve the majority of proposals without edits (extraction quality is calibrated)
- Memory staleness rate drops after several weeks of scheduled runs
Counter-evidence
Scheduled extraction without a review gate becomes unaudited automated memory. Low session volume produces noisy signal with poor pattern quality. The extraction process can surface coincidental patterns rather than causal ones.
Cross-references
- A separate grader agent in its own context window closes the output verification loop at production scale (Anthropic): paired feature — Outcomes closes the output loop, Dreaming closes the memory loop
- A trace alone teaches nothing; learning requires feedback attached to the trace (Harrison Chase): traces without feedback are incomplete; Dreaming applies that principle to persistent state
- Close the feedback loop by mining session transcripts for patterns to promote into config: mining transcripts to promote patterns into config