What We Shipped in July 2026: Make It Accurate, Make It Fast

July 2026

July was a month of tightening things up. We closed out a docs accuracy campaign that took our stale-page count to zero, shaved 200ms off homepage load time, and published three posts — including a build-in-public story about a bug that wasn't the bug we thought it was. Here's what shipped, what we fixed, and what we're seeing in the data.

Docs accuracy: from 44 stale pages to zero, and a generator that keeps it that way

At the start of July we had 44 docs pages that didn't match current CLI behavior — outdated flags, renamed commands, examples that no longer ran as written. By early July, that number was zero.

The easy version of this story is "we fixed 44 pages." The real story is that we didn't stop there. The root cause was that our CLI reference was hand-maintained and drifted from the source every time a command changed. So we rebuilt the generator to derive command groups directly from the CLI source code, with a drift guard that fails the build if docs and source disagree. Code samples across the docs site are now compile-verified against the SDK's shipped types as part of review before a page ships.

This is the same lesson we learned in June with the wrong package name sitting in 61 files for two months: a one-time fix doesn't stay fixed unless the thing that caused the drift gets fixed too. In June that meant find-and-replace across every file. In July it meant building a system that can't drift again.

Try it yourself. Sign up free and run mx setup — the docs you land on match the CLI you just installed.

Homepage performance: LCP down to 2.8s, Lighthouse 100 on desktop

We converted six-plus homepage sections from client-side to server-rendered components, replaced our framer-motion animations with plain CSS, and lazy-loaded the FAQ section instead of shipping it on first paint.

The result: Largest Contentful Paint (LCP — the point where the biggest visible element finishes loading, a core Google ranking signal) went from 3.0s to 2.8s. Lighthouse — Google's page-quality scoring tool — now reports 96 on mobile and a clean 100 on desktop.

None of these were rewrites. Each one was the same homepage, built to load faster: less JavaScript shipped to the browser, fewer things competing for the main thread on first paint, and content that renders progressively instead of all at once.

Three posts, one theme: show the real debugging work

We published three posts in July, and all three are tutorials or build-in-public stories grounded in things we actually ran into.

Wiring Shared Memory Into a Multi-Agent Claude Code Team (July 5) walks through connecting MemNexus to a lead agent and its subagents — named state memories, codeContext tagging, and the concurrent-write gotcha that shows up the first time two agents save at once.

PostHog's Bot Filter Worked Fine. Our QA Check Was the Bot. (July 10) is the one we're proudest of publishing, because it's the one where we were wrong first. We nearly shipped a P0 fix for what looked like a site-wide analytics outage. It turned out PostHog's bot filter was working exactly as designed — it had correctly flagged our own QA check as a bot. We wrote up the real check order and the test we built so it can't happen twice.

Debugging MCP Tool Call Failures: A Field Guide for Memory Servers (today) covers five failure modes for the MemNexus MCP server — server not appearing, auth errors, timeouts, stale tool schemas after an upgrade, empty search results — with the exact command to diagnose each one. We rewrote the docs site's MCP troubleshooting section alongside it, so the post and the docs now say the same thing.

The things we fixed

The bot-filter story above is really a "things we fixed" post on its own, so we won't retell it here — read it for the full account of what we almost shipped and why it was wrong.

The other fix worth naming plainly: the docs drift covered above wasn't a one-off cleanup. Hand-maintained reference docs will drift from a CLI that ships changes weekly. We didn't just close the gap that existed in July — we removed the mechanism that opened it in the first place.

What the traffic data tells us

Google organic traffic roughly doubled from the first week of July to the third. We're not reading too much into a two-week window, but it's the sharpest short-term jump we've seen. Blog posts remain the dominant entry point to the site, consistent with what we saw in June.

On the content side, July continued a pattern from the AI coding tool series: we spent part of the month consolidating and differentiating posts we'd published in a batch as part of the tool-guide series, tightening up content that had overlapping angles on similar topics. We don't have enough post-consolidation data yet to call the result — that's a read for the August changelog.

Both the multi-agent tutorial and the MCP debugging field guide target the same "how do I get this working" search intent that's driven the tool-guide series since February. We'll have a real read on how each is performing in a few weeks.

What's next

August continues the accuracy and speed theme: more framework-specific tutorials in the same hands-on style as July's multi-agent post, continued homepage performance work, and extending the compile-check discipline to tutorial code blocks in blog posts is on the list.

Get started

If July's theme resonated — accurate docs, a fast homepage, and debugging content you can actually use — the best next step is to try the thing the docs and tutorials are about.

npm install -g @memnexus-ai/cli
mx auth login
mx setup

Sign up free at portal.memnexus.ai

---

Read July's posts

• Wiring Shared Memory Into a Multi-Agent Claude Code Team
• PostHog's Bot Filter Worked Fine. Our QA Check Was the Bot.
• Debugging MCP Tool Call Failures: A Field Guide for Memory Servers
• June 2026 changelog