Background memory consolidation

Dreaming keeps multi-session agent memory clean between sessions.

Agents accumulate memory while users work. Over longer trajectories, that memory needs a second pass: deduplicate repeated facts, resolve corrections, compact sprawling notes, and preserve the trail. MemexAI dreaming does that in the background after memory has been quiet.

Why it exists

Memory is trajectory infrastructure, not just recall.

A returning agent does not only need facts. It needs a durable record of what changed, what was corrected, what should be carried forward, and what can be safely ignored next time.

Multi-session memory gets messy

Agent trajectories produce partial decisions, repeated facts, contradictions, and notes that made sense in the moment but need consolidation later.

Context windows are not continuity

A larger prompt can carry more text, but it cannot decide what should survive. Dreaming turns accumulated memory into cleaner state for the next session.

Cleanup still needs an audit trail

Dream writes use normal memory_write and memory_patch paths, so revisions and access logs stay intact instead of hiding consolidation in a black box.

What dreaming solves

It absorbs the first wave of memory health and compaction work.

Duplicate facts across active memory files
Fragmented notes that should become one stable record
Direct contradictions after a user corrects themselves
Low-signal memory that makes future recall noisy
Readable memory files for agents that return days or weeks later
Operator visibility into background consolidation status
Research context

Dreaming is product infrastructure, not a naming gimmick.

Agent memory is moving from simple recall toward maintained state. Anthropic's Managed Agents Dreams is a strong signal for the category. MemexAI applies the same consolidation pressure to app-owned memory: scoped files, normal write paths, revisions, access logs, and operator controls.

The category is moving toward consolidation

Anthropic Managed Agents Dreams turns memory cleanup into a first-class primitive: read past sessions, curate memory, and produce a cleaner store for future work.

MemexAI keeps the source of truth in your stack

Dreaming runs against memory owned by your application, not a hosted agent runtime that owns the memory lifecycle.

The product page is the operator surface

The deeper research argument lives in the blog. This page focuses on what teams can enable, inspect, pause, and operate today.

How it runs

A quiet, bounded cleanup loop for durable memory.

Dreaming does not mine raw transcripts or rewrite memory through a hidden path. It waits for memory to settle, skips logs, writes through normal memory tools, and keeps the trail visible.

1

changed memory files

2

quiet grace period

3

bounded consolidation

4

normal memory patch

5

revision + access log

6

cleaner next session

Operator control

Dreaming is automatic, but not uncontrolled.

Pause globally or per user

Use the global dream config as the master switch, or pause one user from the Dreams panel/API when a workspace needs manual review.

Inspect every effect

Dream writes create revisions with actor dream-agent. If the agent finds nothing useful to update, it records zero files touched and avoids adding noise to dream-log.md.

Keep memory high-signal

The dream agent can merge duplicates, compact fragmented notes, and resolve direct corrections without turning raw session history into permanent user identity.