Once NEXO gained a shared brain across Claude Code, Codex, and Claude Desktop, the next question was obvious: is the parity only true when users stay on the golden path, or is it still true in messy real use? Version 2.6.16 pushes that answer further in the right direction.

At the same time, we also tightened the memory system itself. Retrieval now makes better default choices, memory decay grows more individualized, and Deep Sleep looks farther back before it decides what still matters.

Codex Now Gets Managed Global Bootstrap and Model Sync

Codex sessions launched through nexo chat already started with the current NEXO bootstrap. The weak spot was sessions opened more directly: they could still drift back toward plain Codex behavior depending on how the client loaded global startup state.

NEXO now manages that layer too by writing the active bootstrap and runtime profile into ~/.codex/config.toml. That means Codex inherits:

This still does not magically create Claude Code’s native hook surface inside Codex. But it does make the baseline startup identity far more reliable, even outside nexo chat.

Session Tokens and Followups Are No Longer Claude-Shaped

Some of the surrounding runtime logic still assumed that every interactive session looked like Claude Code internally. That was subtle technical debt: the product claim said “shared brain across clients,” but pieces of coordination still smelled Claude-first.

2.6.16 generalizes that surface. Startup and session registration now support generic external session tokens and client identifiers, and dashboard followups launch the currently configured terminal client instead of hardcoding Claude Code.

HyDE and Spreading Activation Now Make Smarter Default Choices

Two retrieval features were already present in NEXO, but they were effectively opt-in: HyDE query expansion and spreading activation. That meant the capability existed, but the system was still leaving quality on the table unless something called those modes explicitly.

Now the defaults are smarter:

That improves semantic recall without turning every precise lookup into fuzzy mush.

Memory Decay Is Still Ebbinghaus, but Less Generic

We did not replace NEXO with a foreign review scheduler. Instead, we made the existing memory model more individualized.

Each STM and LTM memory now carries two lightweight factors:

Decay still uses Ebbinghaus as the base. These new factors just stop every memory from being treated like the same anonymous unit.

Deep Sleep Now Looks Across a 60-Day Horizon

Before this release, Deep Sleep was already transcript-aware across Claude Code and Codex. What it still lacked was a deliberately longer memory of its own: the ability to mix recent context with older signals instead of focusing almost entirely on the immediate past.

NEXO now builds a blended 60-day context that combines:

The synthesis prompt now explicitly asks for multi-week recurring themes, older/current cross-domain connections, and topics that keep being mentioned but never formalized.

Guardrails Against Sliding Back Into a Claude-Only World

One of the risks in a multi-client system is not the first parity release. It is the third or fourth change afterward, when someone adds a feature and quietly assumes Claude-only paths again.

2.6.16 adds regression audit coverage for exactly that class of mistake: shared automation runner usage, transcript-source parity, Codex bootstrap guidance, and client-aware dashboard followups now have explicit tests.

What This Release Actually Means

Install or Update

New install:

npx nexo-brain

Existing installation:

nexo update

Need to reapply shared-brain wiring and managed client state later?

nexo clients sync

Read the full release history in the changelog or contribute ideas and fixes on GitHub.