The AI memory architecture debate is happening on Twitter, Substack, and YouTube. Both sides are converging toward a hybrid. The live reference implementation of the answer is already operating at sideguysolutions.com.
Andrej Karpathy posted a wiki idea for AI memory. It got 41,000 bookmarks in a week. The OpenBrain debate broke out right after. The disagreement looks like competing approaches but it's actually the same problem solved from opposite directions: where does the AI compile its understanding — at write time, or at query time?
| DIMENSION | KARPATHY WIKI · WRITE-TIME | OPENBRAIN · QUERY-TIME |
|---|---|---|
| Compile when | At write time (compiled understanding) | At query time (synthesis on demand) |
| Mental model | Study guide that's already organized | Filing cabinet with a librarian |
| Wins at | Deep research mode · context loading | Precise structured operations · audit |
| Token cost | Cheap on read · expensive on write | Cheap on write · expensive on every query |
| Risk | Editorial trap — errors baked in forever | Token burn re-deriving same connections |
| Breaks at | Editorial drift over time | Scale + multi-agent access |
| Where it stalls | Wiki stillness (frozen compilation) | Database stillness (no synthesis layer) |
Both sides converge on the hybrid: graph database over structured data + AI as maintainer, not oracle. That's the architectural answer. But neither side has named the calibration layer that decides whether the hybrid actually works at scale.
SideGuy operates all four mechanics simultaneously. Each one is calibrated by an operator-honest doctrine that PJ has been writing into the system for six months. Here's the implementation receipt:
Surfaced from a lead-dev field signal mid-day 2026-05-12: Honcho by Plastic Labs is the open-source proof that the architecture this page names isn't speculative · it's already shipping with 3.4k GitHub stars. Their tagline: "Honcho has defined the Pareto Frontier of Agent Memory."
Their architecture is exactly the hybrid:
What Honcho ships = the general-purpose memory infrastructure (open-source library you embed). What SideGuy operates = the operator-specific compiled doctrine library (curated operational knowledge with editorial discipline). Different layers of the same stack. When SideGuy's doctrine corpus scales beyond markdown files, Honcho-style infrastructure becomes the natural storage/retrieval upgrade path. Until then, MEMORY.md + the auto-link engine works fine — the editorial discipline matters more than the storage layer.
The "missing axis" named below is still missing from Honcho too. The architectural hybrid is solved · the editorial calibration is the next frontier.
Karpathy's wiki risks the editorial trap (errors baked in). OpenBrain risks token-burn (re-deriving the same connections forever). The hybrid solves both — but only if the compilation step has an editorial policy that catches the errors before they bake in.
That's the axis nobody in the debate has named yet. Here's what operator-honest editorial discipline actually requires in code:
Save scars, not summaries
Each doctrine surfaces friction, edge cases, admitted uncertainty. Polished compilation = baked errors. Scar-bearing compilation = correctable signal.
Date every revision
Memory files include dated revision markers. Drift becomes visible. Readers can tell what was true when, not just what's claimed now.
Allow "I don't know yet"
Confidence labels (high · medium · low) are required, not optional. Admitted uncertainty is the operator residue that compounds trust over time.
Anti-slop filter at write time
89-term banned vocabulary catches generic vendor language before it bakes in. Forces compilation to stay specific.
Operator override of pattern-match
When the AI's pattern-match conflicts with the operator's lived experience, the operator wins. Always. The system knows what it knows because PJ shipped it.
AI as maintainer, not oracle
Trilly C extends and maintains the doctrine library. It doesn't claim to BE the answer. The compilation is the operator's; the AI is the editor.
Editorial discipline is the missing axis. Both Karpathy's wiki and OpenBrain's filing cabinet fail without it. With it, the hybrid actually works at scale. That's the doctrine SideGuy has been operating quietly for six months — and the doctrine the debate is going to converge on next, whether or not anyone gives it a name.
When the AI memory debate finishes its laps and the hybrid wins, the next question becomes: who calibrates the compilation step so the editorial trap doesn't poison the wiki? The honest answer requires a kind of operator who has actually shipped, lived, and revised compiled understanding under production pressure. That kind of operator is rare. SideGuy is one.
"I'm almost positive I can help. If I can't, you don't pay."
— PJ · SideGuy Solutions · 858-461-8054 · sms:+18584618054
If you're picking a memory architecture or context layer for production AI, text the line above with one sentence about your stack. Five sentences back from PJ — no calendar link, no demo. Either there's a hybrid pattern that fits or there isn't, and you'll know within an afternoon either way.
"Saved my first SideGuy doctrine in November. Half of the early ones got revised within a month because the lived experience contradicted the compiled summary. Dated revisions saved the system — without them I'd have shipped a wiki that quietly drifted away from the reality I was operating in."
PJ — on why date-every-revision is non-negotiable in compiled understanding
"The anti-slop filter caught my own writing before it caught any AI's. When I wrote 'end-to-end' six times in a megapage, the filter blocked the ship and forced the rewrite. The editorial trap I was protecting the system from turned out to be one I was about to walk into myself."
PJ — on operator-honest editorial discipline catching the operator
"The Compound Map at /operator/compound/ wasn't a feature build. It was an emergent need — once the page graph + entity-cluster-overlap data existed, somebody had to visualize the constellation so I could see what was compounding. The graph layer makes the wiki layer legible."
PJ — on how the graph layer surfaces what compiled understanding alone hides
"Trilly C maintains and extends the doctrine library every session. I write the scar; Trilly C writes the doctrine file. The AI is the editor, never the source. When the AI tries to summarize me into something polished, I push back and the compilation stays operator-honest."
PJ — on AI as maintainer not oracle
"The Daily Briefing reads the compiled doctrines AND the live data each morning. Write-time and query-time both running. The interpretation surfaces the scars I'd otherwise have to dig for. That's the hybrid in production — not a thought experiment."
PJ — on why hybrid memory needs both layers running simultaneously, not switching between them
Text PJ ⚡
858-461-8054
Text PJ
858-461-8054
Don't see what you were looking for?
Text PJ a sentence about what you actually need — I'll build you a free custom shareable on the house. No email, no funnel, no SOW.
📲 Text PJ — free shareable