Deep Sleep already knew how to spot recurring errors, protocol drift, abandoned work, and cross-session patterns. The problem was what happened next: it could still create duplicate followups with different IDs, pile overlapping learnings on top of each other, and stop at diagnosis when the real value was a fix.
Followups Are Now Deduplicated by Meaning
Before 2.6.21, Deep Sleep relied mostly on generated IDs and action dedupe keys. That prevented identical replays, but it did not stop semantically identical followups from reappearing with different descriptions.
Now the apply layer checks a proposed Deep Sleep followup against open followups by meaning before creating anything new. If the same issue is already open, NEXO links to the existing followup instead of growing noise. And if the new nightly suggestion is more concrete than the old wording, it upgrades the existing followup in place.
Learnings Now Consolidate Instead of Just Accumulating
Nightly learning growth is only useful if the learning store stays clean. 2.6.21 adds a local consolidation step with four outcomes:
- Duplicate: the learning already exists, so Deep Sleep reaffirms it instead of creating another row.
- Reinforcement: the new learning strengthens an existing one, so NEXO updates the canonical record instead of duplicating it.
- Contradiction: the new learning conflicts with an existing rule, so NEXO creates an explicit review followup.
- Genuinely new: only then is a new learning created.
The key design choice is that this logic lives in the apply phase, not only in the model prompt. That means Deep Sleep gets stronger even when the model output is imperfect.
Recurring Problems Now Backfill Concrete Engineering Work
The third gap was the most important one: Deep Sleep could tell you something was wrong, but still leave you with a passive observation instead of an actionable fix.
2.6.21 pushes the synthesis contract toward explicit fix artifacts. Medium and high-severity recurring patterns are now expected to carry a concrete proposed fix: script, hook, checklist, validation step, workflow change, or guardrail. And if that fix exists in the pattern analysis but no actionable followup was emitted, NEXO backfills one automatically.
That changes the character of the overnight loop:
- less duplicate followup churn
- stronger canonical learnings
- explicit contradiction handling
- recurring problems turned into concrete engineering work
Why This Matters
If you want an agent to become materially better over time, nightly analysis cannot just write more memory. It has to improve the shape of memory and push the system toward durable fixes.
That is the real point of this release: Deep Sleep is still an analyst, but it is now much closer to being an engineer too.
Install or Update
New install:
npx nexo-brainExisting installation:
nexo updateSee the full version history in the changelog, or help drive the next strategic step through the issue that tracked this work.