Dream pass

Inspect Kairos’s offline consolidation pipeline.

Periodically, Kairos pauses ordinary work and consolidates memory — a structured offline pass we call a dream. It is structurally analogous to sleep-dependent memory consolidation: sample related observations, find non-obvious connections, distill durable lessons, stage the result for the agent’s next wake-time review.

Dreams do not mutate long-term memory directly. They produce candidates that the agent accepts or rejects on the next wake. You always have an opportunity to review before consolidations become durable.

When dreams fire

Dreams are triggered by a pressure score that combines several operational signals (work elapsed since last dream, unresolved retries, observation novelty, log volume, a few more). When pressure crosses a threshold, the next turn is a dream.

A safety floor guarantees a dream every N turns even when pressure is low; an anti-thrash floor prevents back-to-back dreams. Thresholds are tuned ongoing.

backant dream stats

Summary of recent dreams:

backant dream stats

For each recent dream:

• When it happened
• How many candidate consolidations were produced
• How many edge proposals were made
• How many were accepted vs. rejected during the next wake-time review

Use this to verify the consolidation pipeline is actually doing work. A long stretch with no dreams plus lots of unresolved retries is a sign that pressure signals aren’t firing correctly.

backant dream replay

Walk through a past dream as an annotated timeline:

backant dream replay dr_20260515_0a1b

Each event in the dream is logged with its inputs and outputs:

• Sample: which entries were drawn for this dream
• Coactivate: which pairs the model identified as related
• Reflect: how the model rewrote / clustered the related entries
• Bucket-write: which staged candidates were produced
• Edge-propose: which links between entries were proposed

Useful for debugging odd consolidations, or for understanding why a particular edge was proposed.

The five stages (briefly)

1. Slow-wave-style curation        — distill failure signatures from recent logs
2. Candidate sampling              — pick pairs of entries to consider together
3. Co-activation                   — does this pair belong together?
4. Reflection                      — what should the merged / linked form look like?
5. Bucket-write + edge-propose     — stage candidates for the agent's review

The pipeline is deliberately advisory. Anything that lands in durable memory has been through the agent’s verdict at wake time.

Reviewing dream output

You don’t normally need to review dream output by hand — Kairos does that automatically on the next wake via the wake_edge_triage and dream-bucket verdict tools. But if you want to:

backant memory stats              # see dream_bucket_pending count
backant dream replay <dream-id>   # see what was produced

If a recent dream produced consolidations you disagree with, the next wake-time review will reject them based on the agent’s read of the codebase. The pipeline is self-correcting in that sense.

Why offline?

Mutating memory on every write would couple every observation to a reasoning pass — expensive and noisy. Batching consolidation into structured offline passes:

• Keeps live work fast (no consolidation cost paid on the critical path)
• Lets the consolidation use a stronger reasoning model than live work
• Mirrors how biological consolidation actually works during sleep

The architecture page goes a little deeper on the design intent.