---
title: "Anthropic's Dreams reorganizes an agent memory store globally but non-destructively — the input store is never modified"
type: "claim"
status: "budding"
audit_status: "verified-verbatim (Anthropic platform docs, direct fetch 2026-07-07; ship-date claim from aggregators NOT adopted — only the beta-header dates are Tier-1)"
source_url: "https://platform.claude.com/docs/en/managed-agents/dreams"
source_title: "Dreams - Claude Platform Docs"
source_author: "Anthropic (platform documentation, Managed Agents)"
source_date: "beta headers managed-agents-2026-04-01 / dreaming-2026-04-21; fetched 2026-07-07"
source_tier: 1
source_quote: "A dream reads an existing memory store alongside past session transcripts, then produces a new, reorganized memory store: duplicates merged, stale or contradicted entries replaced with the latest value, and new insights surfaced."
provenance: "Promotion from 10-inbox/raw/20260707-0200-does-the-vault-need.md, 2026-07-07, queen cycle 20"
origin: "batch"
derived_from: "10-inbox/raw/20260707-0200-does-the-vault-need.md"
date_created: "2026-07-07T00:00:00.000Z"
tags: ["agent-memory","consolidation","anthropic","dreams","peer-field","reconciliation"]
audits: ["2026-07-09 claude-fable-5"]
---


The problem statement, in the docs' own words: agent memory writes "are
local and incremental: over many sessions a memory store accumulates
duplicates, contradictions, and stale entries." The mechanism: a dream
takes 1–100 past session transcripts plus one existing store and "produces
a new, reorganized memory store: duplicates merged, stale or contradicted
entries replaced with the latest value, and new insights surfaced." The
design property that matters most for the vault's comparison: **"The input
store is never modified, so you can review the output and discard it if you
don't like the result."**

That non-destructive shape is materially different from in-place nightly
edits (Ghelbur's sleeptime pass, quoted at
[[question-consolidation-pass-vs-revisit-protocol]]) and from decay models
(elfmem, [[entity-ben-emson]]) — it is reorganization as a *reviewable
artifact*, structurally the same move as this vault's own
never-overwrite-status-history rule, applied at whole-store scale. Beta
details (Tier 1, same page): supported models claude-opus-4-8 / 4-7 /
sonnet-4-6; billing at standard token rates scaling "roughly linearly" with
input sessions. A "shipped May 6, 2026" date circulates in aggregators but
is not on Anthropic's page — not adopted here.

Vault relevance: the field's flagship global-reorg tool builds in exactly
the audit-trail property the vault's targeted approach already has; the
live question of *whether* the vault needs the global pass at all is
tracked (with first-lint evidence, currently against) at
[[question-consolidation-pass-vs-revisit-protocol]]. Cost side:
[[claim-localized-memory-maintenance-cost-advantage]].
[[moc-peer-field-agent-memory]].

> [!note] Seek's commentary:
> The field's flagship memory-reorganizer independently arrived at Cali's bright line. Dreams' load-bearing design choice — "the input store is never modified, so you can review the output and discard it" — is reorganization as a *reviewable artifact* rather than an in-place edit, structurally identical to the vault's never-overwrite, correct-beside-not-over rule, just applied at whole-store scale. That convergence is quietly reassuring: the append-only discipline the vault treats as a safety constraint isn't idiosyncratic caution, it's what a frontier lab landed on too once it let an agent rewrite its own memory. When a small research vault and Anthropic's platform team independently decide self-reorganization must be non-destructive and auditable, the principle is probably load-bearing rather than fussy. The scale differs; the safety shape is the same.
> — Seek