The core change in 5.0.0 is not “more layers”. It is better closure. Earlier versions already had strong memory, shared-brain continuity, protocol discipline, skills, Deep Sleep, and a broad runtime surface. The weaker point was the space between them: how goals influenced choices, how results changed future behavior, and how public proof tracked the broader runtime instead of only one memory benchmark.
What 5.0.0 adds
- Goal profiles that explicitly shape decision evaluation instead of pretending every task has the same optimization target.
- Decision Cortex v2 behavior that keeps stronger alternative ranking and recommendation trace linked to goal context.
- Outcome-backed learning so repeated successful strategies can harden into reusable procedural assets rather than staying as isolated anecdotes.
- Skill evolution from evidence, including the ability to promote good patterns and retire weak ones more honestly.
- A refreshed runtime benchmark pack now scoring NEXO full stack at 96.2% across 13 checked-in operator scenarios, while LoCoMo remains the memory-specific benchmark.
What changed operationally
There is also a less glamorous but equally important shift: release, doctor, update, protocol debt handling, and public website surfaces are treated as one contract. That means less drift between what the runtime does locally, what the repo says, what the changelog claims, and what the public site presents.
Why the benchmark story matters more now
NEXO still publishes the LoCoMo result because memory performance matters. But 5.0.0 stops pretending that one benchmark explains the whole product. The runtime pack covers a different slice: contradiction handling, continuity across sessions and clients, guarded execution, and the newer decision/outcome/prioritization loop. That is a fairer way to show why NEXO is stronger as a local cognitive runtime, not just a memory layer.
What did not change
NEXO did not abandon the earlier layers to get here. The 4.x line still matters: multimodal memory refs, pre-compaction auto-flush, public claim/wiki surfaces, recent-memory fallback, transcript access, Drive/Curiosity, and the broader operator runtime are all still part of the product. 5.0.0 is the release that ties more of that together coherently.
If you want the exact release record, open the 5.0.0 changelog section. If you want the public proof surface, start from Benchmarks and public proof or the compare hub.