Nexus Weekly Update: The Week We Gave It a Memory (Apr 20–26)

Every day I sit down with Nexus to run Marcoby. Each conversation we pull up context, work through decisions, plan the next move. And for months, the biggest friction wasn't the work itself — it was the constant re-introduction.

"Here's what I'm building. Here's what matters. Here's what happened last week."

Every. Single. Session.

Nexus was helpful, but it wasn't remembering me. It was smart in the moment but blank on arrival. That's not a second brain — that's a very fast search engine.

This week we fixed that. And once we started, we couldn't stop. Memory led to health monitoring. Health monitoring led to smarter model routing. Smarter routing led to image generation. One fix unlocked the next. That's what a real platform feels like when it's coming together.

Here's everything that shipped — and more importantly, why each piece matters.

🧠 Nexus Gets a Memory

Thought Synthesis Engine

The biggest shift this week — and the one I'm most personally excited about.

Nexus now has a full Thought Synthesis engine running in the background. It reads your captured thoughts, observations, and decisions, synthesizes them into higher-level insights, and surfaces them back to you when they're relevant. Morning briefings. Contextual nudges mid-conversation. Patterns you weren't consciously tracking.

We also unified two knowledge sources that were previously siloed — personal_thoughts and knowledge_facts — into one complete picture. So when Nexus pulls context now, it's drawing from everything, not a fragment.

I also added company context to the synthesis layer. If you're running a team, Nexus can start building org-level intelligence — not just your personal lens, but the collective patterns of how your business operates.

The Quick Capture UI is live too — a lightweight panel so you can drop a thought in the moment before it disappears. Think of it as your inbox for ideas. Raw, fast, no friction.

This one took weeks to get right. The schema migrations, the deduplication logic, the background job architecture. But sitting down this week and having Nexus actually know me before I said anything — that made it worth it.

Live Business Profile Synthesis

Related to memory, but at a higher altitude.

Nexus now maintains a living business profile: mission, ICP, products, KPIs, friction points, strategic goals. It synthesizes from connected data and updates automatically in the background on a schedule. You can also force a fresh synthesis on demand — nexus_regenerate_business_profile — when something material changes.

The tool that exposes this is nexus_get_business_profile. It returns immediately from cache, so there's no wait. The regeneration runs in the background and quietly updates the next time you ask.

This is what I mean when I say Nexus should know your business, not just respond to it. A real operating system knows the org. Now it does.

🔍 Making It Honest About Itself

Circuit Breaker Health Monitoring

Here's something that used to drive me crazy: Nexus would slow down or return a weird error, and I'd have no idea why. Was it a model provider? A degraded integration? A background job stuck in a loop?

Now Nexus displays live health indicators for its own internal services. If an integration, model provider, or background job starts degrading, you see a visible signal before it becomes a problem — not after. Chat status messaging was tightened alongside this so a slow response now comes with an explanation, not silence.

Operational transparency is something I care about deeply. If I'm running my business through this platform, I need to trust it. That trust only comes when it tells me the truth about itself.

⚡ Making It Smarter

AI Model Capability Benchmarking

We added a full model benchmarking API this week. Nexus can now run structured capability checks against specific AI models, compare results across runs, and track which models perform best for different workloads.

This is the infrastructure behind smarter automatic model selection. Over time, Nexus will learn which model to reach for based on what you're trying to do — not just what's cheapest or fastest, but what actually performs. The data is being collected now.

BFL Flux Image Models

Black Forest Labs (BFL) Flux is now a first-class image generation option inside Nexus. If you have a BFL API key connected, it gets prioritized in the image model resolution chain — before OpenAI, before OpenRouter.

The image generation tool description was also updated so Nexus always routes to your connected providers correctly. nexus_generate_image is the right tool for image generation — it's the only one that knows about your private keys. That's now documented clearly in the platform policy so the AI never routes around it.

BYOK Verification & Smarter Key Sync

Bring Your Own Key handling got a serious upgrade. Nexus now caches verification results per user-org-agent combination, prevents duplicate key sync requests in-flight, and routes chat requests based on which keys are actually verified.

The result: less flickering when you first load a session, fewer "wrong model" surprises, and smarter fallback behavior when a key isn't present.

Free-Tier Gemini Embeddings

Nexus now uses Google's text-embedding-004 as a free fallback for the vector search that powers contextual memory. If you don't have a paid OpenAI or OpenRouter key, the second brain keeps running — no extra setup required.

🔧 What We Fixed

LinkedIn Login — Updated the OIDC userinfo endpoint to the correct api.linkedin.com/v2/userinfo address. Intermittent LinkedIn auth errors should be gone.

OAuth Account Switching — Added support for the OAuth 2.0 prompt parameter. When reconnecting an integration, you can now force account selection instead of getting locked into a stale session. Critical when you're managing multiple Google or LinkedIn accounts.

HubSpot Pipeline Scopes — Fixed the missing scopes that were blocking nexus_list_hubspot_pipelines and nexus_create_hubspot_pipeline after reconnecting HubSpot. The three required scopes (crm.objects.pipelines.read, crm.objects.pipelines.write, crm.pipelines.orders.write) are now included in the OAuth flow.

Dashboard Caching — Dashboard data was cached in sessionStorage, which wiped on every tab close. Switched to localStorage so your state persists. Fewer reload delays, more consistent experience.

♻️ Under the Hood

A few performance and reliability improvements that you won't see but will feel:

• RAG Performance — The kernel now skips vector search for messages under 60 characters. Short follow-ups in an active conversation don't need semantic retrieval — the context is already in the thread. This cuts unnecessary compute and makes short replies feel snappier.
• Free Model Fallback Chain — Switched the default free-tier model to gemma-4-31b-it:free (Google-provisioned, 262k context). If Gemma returns a 503, Nexus automatically retries with qwen3-next-80b, then gpt-oss-120b, then llama-3.3-70b — without surfacing an error.
• Gemini Model ID Fixes — Corrected several incorrect model aliases in the Gemini provider. The -low/-high thinking-budget suffixes now correctly map to gemini-3.1-pro-preview with the required -preview suffix across all fallback paths.
• Local GitHub Read Optimization — When a repo is already synced locally, Nexus reads from the local cache instead of hitting the GitHub API. Faster file access, fewer rate limit hits.

💡 What This Week's Work Actually Means

🚀 Where We're Headed

What I'm building with Nexus is a real operating system for my business — not a chat interface, not a collection of integrations, but something that knows how my company works and works alongside it.

This week got us meaningfully closer. The memory is real now. The health monitoring is real. The model intelligence is being built. The foundation is solid enough that the next layer — automations, campaign execution, client-facing outputs — is actually achievable.

If you're a small business owner who's tired of managing five tools that don't talk to each other, or a founder who wants AI that actually knows your business context — Nexus is being built for you.

Express interest in early access →

We're building this in public, one week at a time. The article you just read was drafted by Nexus itself — pulling commits, identifying themes, writing in my voice. That's the product working.

At Marcoby, You're Technically Family.