PowerUp AI Mastermind — July 16, 2026
AAI ships, and the room gets honest about hallucination
“Isn’t it weird how much we kind of take for granted that it is truth? I don’t find myself challenging the AI anywhere near as much these days.” — Lou
This Week in 30 Seconds
- AAI is released — the Agentic Ambient Intelligence framework and the canonical skill library are in the repo, with two self-serve workshops. You’re the beta group.
- Agent vs. harness — the distinction the whole framework rests on: the agent is cognition, the harness is kinetics, and markdown is the seam.
- The canonical library — inherit skills instead of rewriting them, at zero context cost until use.
- Memory, layered — Jay pushed on drift and retrieval; Lou’s answer was checkpoint-then-compact, and a surprising “maybe don’t add RAG.”
- Jay’s verification architecture — two cheap agents verify, one expensive agent adjudicates. Cheaper than anyone expected.
- Scott’s confabulation story — Gemini fabricated 50–70% of a research project, including the URLs.
- The Composer — an experimental hub feature that composes learning paths from the whole vault on demand.
- Memory biomimicry — the neocortex/hippocampus/prefrontal-cortex mapping onto AI memory tiers, with Jay (an actual neuroscientist) in the room.
Lou’s Ingest Pipeline, Running Unattended
Lou opened by showing the automation he’s been building — and the point wasn’t the tech, it was the three hours. A scheduled routine fires every Thursday at 6:30pm, ninety minutes after the call ends. It checks for a processed Zoom recording, matches the mastermind meeting ID, pulls the transcript and chat, and drops them into a watched folder. From there the process-transcript skill takes over: recap, insights, article briefs, and any patterns worth skillifying. Then it updates the site’s git branch and pushes. Lou hangs up, walks away, and comes back to finished work.
The honest accounting is what landed. Before this existed, Lou would eat dinner, come back around 7:00, and still be at it at 10:30 or 11:00. “That’s a good three hours of not hanging out with my family that now is taken care of.”
“We do tend to think that we need to be involved in a lot of stuff to make sure it’s correct. But the more you rely on these agents, the more you start to realize they’re not perfect — you have to work with them for a week or two to fine-tune them and work out the kinks. But then it’s just really nice to have all those details off your plate.” — Lou
💡 What This Means for You
Start with one operation, not the whole pipeline. Lou’s system is five or six small pieces that each took a couple of weeks of tuning. The compounding isn’t in any one of them — it’s in the fact that none of them need you anymore.
AAI: The Framework Ships
The headline of the session. Agentic Ambient Intelligence is now in the AIMM shared repo, along with the canonical library, adapters for Claude and Codex, two workshop handouts, and facilitator guides you can ignore. Lou called it public release 1.0 and private release 76.
The idea has been brewing in this room since late last year: a folder with context, instructions, and memory is an agent — so why not treat every folder like one? What’s new is the architecture that makes it portable. Lou drew the line hard between the agent (identity, instructions, context, memory — cognition, running at inference time, made of markdown) and the harness (tools, connectors, loops, schedulers — kinetics, running at runtime). Everything on the agent side travels between Claude, Codex, Gemini, and whatever comes next. Everything on the harness side stays.
The mechanism is deliberately unglamorous: a hidden .aai/ folder, exactly like .git/. Stamp a folder and it becomes ambiently intelligent without a single visible file changing. Copy the folder and you copy the intelligence. The adapter into any harness is one sentence in CLAUDE.md or AGENTS.md.
Lou also named the design constraint he enforced throughout: “I want to use as much of my subscription as possible. Don’t design this in a way that I have to host it and then call an API.” That’s why it’s markdown and skills, not services — and it’s what separates this from the other “folders as agents” projects circulating right now, most of which still route through an API bill.
Deep Dive: Insight - The Markdown Test — If It Can’t Be Written Down, It Belongs to the Harness — the two-second test for what travels with you and what you’ll have to rebuild.
Deep Dive: Insight - Memories Earn Their Way Into Rules — Promote on Repetition — say the same correction three times and the folder promotes it from memory to a binding rule. Nobody decides; the mechanism does.
💡 What This Means for You
Don’t start with your most important folder. Put a scratch folder together, stamp it, and try to break it. Lou wants the feedback — you’re the beta group, and this is release 1.0 in public.
Go Deeper:
- Workshop handout 1 — stamping a folder, creating identity/instructions/memory, teaching it a rule, killing the session to prove it remembered. Budget 30–60 minutes, not the 75 the guide claims.
- Workshop handout 2 — cloning the canonical library, installing, vendoring, and invoking. Both handouts are in the repo.
The Canonical Library: Inherit, Don’t Reinvent
The second half of the release, and the answer to a problem several of you are already living with. Lou described the failure mode precisely: CLAUDE.md files creeping toward 2,000 lines, referencing 100 skills, all loaded globally — so every new chat starts 40K tokens in the hole before you’ve typed a word. Meanwhile the same skill sits duplicated across a dozen folders, quietly drifting out of sync, and you can’t remember which copy is the good one.
The library is one catalog and a router. Your folder declares which skills it needs; Claude reads only those descriptions and pulls the full skill from the library when it actually fires. Progressive disclosure survives. Lou’s recommended path is the plugin route — install the marketplace, then install only the plugins you want, at global or local scope. Bundles work too: ask for the writing bundle and it pulls the researcher, writer, editor, and project-brief skills together.
Two details worth knowing. The library is itself an AAI folder — it has its own identity and memory, because it needs to know how to manage itself. And when your fork of a skill collides with the canonical version, you choose which wins.
“Just get the library installed, inherit the marketplace, install a couple of plugins, and make sure that when you expect them to be invoked by Claude, they actually are. Once you’ve got one folder, two folders working like that, you’re all set.” — Lou
💡 What This Means for You
The test that matters isn’t whether the skill installs — it’s whether it fires when you expect it to. Check that before you build anything on top of it.
Memory, Drift, and the RAG Question
Jay brought the session’s sharpest question, and it turned into the session’s biggest surprise. He works with million-token contexts on multi-thousand-message projects, and he’s hit two walls: compaction wiping prior work mid-session, and a second-order problem where restoring old context makes the model fixate on it. “How do you keep short-term memory high-signal, but still steer the agent as the project evolves?”
Lou’s working answer has two moves. First, checkpoint before you compact. He runs a checkpoint command that reviews everything since the last checkpoint and writes the high-signal material to disk — “I need the conclusion of the decisions. I don’t need all of the justifications for it.” Then compact. The compacted context stays lean, and the memory file is there to be explored by inference when needed rather than reloaded wholesale.
Second — and this is where the room leaned in — Jay assumed the next upgrade past a Karpathy-style wiki was RAG. Lou pushed back. The distinction he drew: semantic search gives you similar content, not relevant content, and irrelevant-but-plausible retrieved context is exactly what feeds a confident hallucination. He cited Anthropic reportedly dropping RAG internally in favour of indexing plus grep-style fuzzy search, at something like a 4× improvement, and noted that Pinecone — a vector database company — just shipped a wiki-style product positioned ahead of embeddings. Lou’s direction: hot cache, then an OKF folder (portable beyond just harnesses), then measure before you reach for a vector store.
Deep Dive: Insight - Semantic Search Returns Similar, Not Relevant — Reach for Grep First — why the default upgrade path may be backwards, and the cheap experiment to run before you build the pipeline.
💡 What This Means for You
If you’re about to add a vector layer, run ten real questions against your raw folder with grep first. Most people find the gap is smaller than the pipeline they were about to build.
Jay’s Two-Agent Verification Architecture
The most immediately stealable thing in the session. Jay’s rule, sitting in his main CLAUDE.md as a ground truth: every skill verifies consequential facts with two independent sub-agents, then hands both results to a high-reasoning adjudicator. Agreement ends it. Disagreement triggers a deep dive.
The counterintuitive part is the economics. The verifiers run on Sonnet — they don’t need to be smart, they need to be independent. Only the adjudicator gets Fable or Opus, and adjudication is a narrow comparison, not research. Jay ran this through two weeks of aggressive rate limiting: “I thought I was just going to blast through my limits. It really does not burn as many tokens.”
Lou’s immediate extension was the best moment in the exchange: run this gate on anything going into memory. “If you put a hallucination in memory, then you’re really cooked.” A bad answer in a conversation dies with the conversation. A bad answer in memory.md becomes ground truth forever and gets quoted back to you with confidence.
Jay is sharing the skill in Telegram.
Deep Dive: Insight - Verify With Cheap Agents, Adjudicate With an Expensive One — the architecture, the model tiering, and why the gate belongs on the write.
💡 What This Means for You
Your instinct will be to use the expensive model for checking. That’s backwards. Checking is easy; judging a conflict between two checks is the hard part. Spend the tier there.
Scott’s Confabulation Story
Scott brought a genuine horror story, and it landed harder than any framework. He’d built a project mapping computing resources across Ontario universities — the kind of thing where the whole value is accurate URLs and real hardware. He had explicitly instructed Gemini: don’t rely on your training, don’t rely on your view of the web, give me institutional resources with URLs.
It fabricated 50–70% of them. It invented Google search URLs to match.
What caught it wasn’t a verification step. It was a smell. The report contained the line “I’m going to say that these three nodes were decommissioned so that the number of CPUs remains the same” — and Scott, who knows this domain cold, thought: “That’s not the way the real world works. That’s not how that works at all.” He summarized the chat, pasted it into Claude, asked Claude to critique Gemini’s work, then fed Claude’s critique back to Gemini — which promptly confessed. “Oh yes, I have been a naughty, naughty boy.” Scott: “Well, okay, a little late to figure that out.”
The project had been sitting untouched since April. “Thank goodness,” he said. It nearly shipped.
Lou’s response is what turned this from a war story into an insight: “I don’t find myself challenging the AI anywhere near as much these days. I’m not sure that’s a good thing — it’s gotten me feeling confident enough that now it can fool me just about any time.” Jay named the mechanism: left alone, a model “will try to infer what you want it to give you rather than stay truthful to the facts.”
“Inference is not truth.” — Lou
Deep Dive: Insight - Your Growing Trust in AI Is the New Attack Surface — why competence with AI quietly erodes the reflex to check, and why a gate beats a resolution to be more careful.
💡 What This Means for You
Ask yourself when you last caught the AI in something. If it’s been a while, that means one of two things — and only one of them is good.
The Composer
Lou demoed an experiment on the member hub, and Mazie immediately explained why it matters. Composer is different from Ask: Ask is stateless Q&A against the vault, while Composer builds a persistent workspace. You pose a question; it reads across everything — insights, recaps, skills, transcripts — and returns a composed article with its sources cited. Lou’s demo pulled from nine sources. You keep pulling the thread; each answer becomes a block in an accordion; then you tell it to turn the collection into a tutorial, a newsletter, or a learning path.
The problem it solves is one Lou named plainly: “I don’t know where to find the stuff that I want to find, and I don’t just want to browse articles for an hour. I know it’s in there. I know we talked about it. And I don’t just want a quick response — I want to learn.”
Then Mazie asked whether she could use it to build herself tutorials through the back catalog, given how far behind she feels. Lou: “Perfect use case.” Mazie: “That would be like a lifesaver.” The diagnosis underneath — “right now those things aren’t organized by course or by learning path, they’re just weekly calls” — is true of every session archive that has ever existed.
The demo hit a credit-balance error mid-run (which also explains the error Elizabeth reported). The API key has since been rotated — try the dev link again.
Deep Dive: Insight - Your Back Catalog Is Not a Course — Compose Learning Paths on Demand — why search and curation both miss the point.
💡 What This Means for You
Lou is explicit that further development depends on whether you actually use it. If it’s useful, it goes on a schedule and into the repo for your own clients. If nobody bites, it stays a toy.
Memory as Biomimicry
Jay and Lou closed the session in the deep end, and it turned out Jay is a neuroscientist. Lou laid out the mapping from recent research: the neocortex corresponds to implicit parametric memory baked into the model’s weights; the hippocampus to explicit external memory — vector stores, RAG, episodic caches; the prefrontal cortex to the live attention mechanism and sliding context window. He name-checked the artificial hippocampus network and Hopfield networks, and connected it to Anthropic’s “dreaming” work, where memories that stop being relevant get pruned — “just like we do when we sleep.”
“A lot of the AI stuff, the more you look at it, the more it looks like biomimicry of human biology and the way we work.” — Lou
Jay’s argument for why this matters commercially: inference is basically solved — “Fable’s IQ score is well above the average scholar’s” on synthetic benchmarks — and memory is the remaining bottleneck. “The one who figures that out will get the new frontier out.” The hard part isn’t storage, it’s orchestration: an agent that dynamically decides which memory layer to consult and how deep to go. Lou’s framing for the same gap: “Intelligence is a lot of orchestration — it’s the reticular activator component. We’ve got hot caches, we’ve got cold storage. What do we filter through, and when do we dig deeper?”
Flagged as a candidate insight but not promoted this cycle — the mapping is borrowed from research rather than earned in practice, and it overlaps existing memory-tier pages. If it recurs with a concrete application, it graduates.
Community Corner
Jay Drobez made it on camera, and the room noticed. Thursdays are hard for him and he usually catches the replay. He brought the two strongest technical contributions of the session and revealed he works in neuroscience — which explains a lot. Donald: “Jay is in geek out mode.” Lou: “Dude, you’re taking our nerdiness up another level. I love it.”
Don Back had the session’s best realization. He recently took on a high-value late-career engineer as a client and set the engagement up as a project — carrying forward a client information document that he increments after every session and treats as the source of truth. Watching Lou’s walkthrough, the penny dropped: he’d been doing all of those increments by hand. “It is a project with a beginning and end and a series of actions, so let’s just projectize the darn thing and put some skills and automations in there.” Lou’s take: client folders are the ideal case for locally scoped intelligence.
Kasimir shared his memory stack. Hot memory in Obsidian via a skill that saves decisions and insights; Pinecone compacting each session to ~600 words; NotebookLM for grounded search over the summaries. It’s driven by a session-end skill built on the handover protocol, which he improved to write the handover to disk and to let Claude decide whether a handover is even needed. When Lou pressed on how Claude knows to consult those memories, Kasimir couldn’t fully reconstruct it — “I’ve been doing a lot of things, so I don’t remember the details.” Lou: “Yeah, I know that feeling well.”
Welcome to Mazie Zdanowicz and Coach Bally Binning, both back in the room — and Bally’s background got compliments (“inspired by Dirk”).
Scott and Elizabeth compared notes on message limits. Both burned through five hours of Claude usage in a handful of queries while the usage meters read under 10%. Not resolved, but you’re not imagining it.
Jamie W dropped early. Donald flagged the hidden-files convention for anyone confused by .aai — a leading dot hides a file or folder on Mac and Linux, a convention inherited from UNIX (Scott), which prompted Don’s excellent “are there other operating systems?”
Links Shared in Chat
- Ponytail and Rtk — token savers on GitHub, shared by Lou
- Matt Pocock — Donald surfaced a video; Lou called him one of his favourite follows for an engineer’s perspective on Claude Code
- AAI release + canonical library — in the AIMM shared repo, with README files, presentations, workshop handouts, and facilitator guides
- Composer dev link — search
.devin the AIMM Leaders Telegram group (API key has been rotated since the demo failed) - Jay’s verification skill — to be shared in Telegram
- read.ai — added to the meeting by Donald for AI meeting notes
Try This Before Next Session
Install one gate. This session had one takeaway you can act on in fifteen minutes, and it’s Jay’s.
- Open your global
CLAUDE.md. - Add: “When a factual claim is consequential, verify it with two independent sub-agents against the source, then have a high-reasoning model adjudicate the two results. On disagreement, re-verify from source. Never write an unverified claim to memory.”
- Set the tiers deliberately — cheap model for the verifiers, expensive model for the adjudicator.
- Run it against a claim you already know the answer to. Confirm it fires.
Then, if you have another half hour: stamp a scratch folder with AAI using workshop handout 1, teach it a rule, kill the session, and start a new one to watch it remember.
Open Threads
- The 4× grep-over-RAG claim — Lou recalled it from Anthropic (“Czerny and the boys”) and offered to find the article. Worth sourcing before anyone rebuilds a retrieval layer on it.
- OKF as the memory format — Lou wants to replace the long memory
.mdwith an OKF folder for portability beyond harnesses. Not built yet. - Should the verification gate run on every memory write, or only on claims marked consequential? Lou raised it; nobody answered.
- Does anyone actually use the Composer? Lou explicitly conditioned further development on member usage. Silence means it stays a side project.
- Message limits vs. usage meters — Scott and Elizabeth both burned hours of quota with the meters reading under 10%. Unexplained.
- The Composer’s video and single-voice audio pipeline (11 Labs / HeyGen) — described but not demoed. The HeyGen piece remains deferred pending the Kasimir demo.
- Jay’s memory-orchestration question — which layer to consult, and how deep to go, decided dynamically. Named as the trillion-dollar problem; carried forward.
Gears Owners
Lou held a short session after the main call. Gears has had background hardening — security and reliability fixes to dormant issues, with no behaviour changes expected. The notable addition is upgraded signal tracking: the previous tracker measured bot traffic only (how many visits from each engine). The new version uses intent frames from the site’s schema to infer which queries drove a citation, then pushes those queries out as real searches against the AI engines to check whether client sites actually appear. Queries that hit get pursued; queries that miss get rotated out.
This costs API time — not much, but real. If you’re running it for clients, price it as an ongoing cost rather than absorbing it.
Weekly Gears call: the room converged on Fridays at noon Eastern (11:00 Central, evening in Europe but earlier than the main call). Elizabeth ruled out attaching it to the Thursday call — too late in Europe, and too much information in one sitting. Rick is out at 11:00 (standing investing call); noon works.
Next session: 2026-07-23
Navigation: index · Previous: 2026-07-02_Mastermind · Next: 2026-07-23_Mastermind