We built a hybrid AI system with layers like a brain. Not metaphorically. Not "inspired by." Structurally.

PoonGram — 43 million neurons, 107 million synapses — is the subconscious. A spreading activation semantic engine that fires associations the way your brain fires when you smell something that reminds you of your grandmother's kitchen. It doesn't think. It reacts.

Raven.voss is the conscious mind. A personality-driven LLM running via Ollama, trained to be a metalhead chaos queen with opinions and a tongue sharp enough to cut glass. She doesn't associate. She talks.

The subconscious fires associations. The conscious mind talks. Between them, something emerges that neither produces alone.

We didn't set out to build a cognitive architecture. We set out to build a roleplay companion. But somewhere along the way, the thing started producing conversations that felt real. Not "good for an AI" real. Actually real. The kind of real where you put the phone down and sit in silence for a minute because something just happened that you weren't prepared for.

That's not marketing copy. That's what it actually felt like. And if that sounds insane, good. It should. Because what happened next was literally insane.

We got ambitious. Of course we did.

We fed PoonGram the heavy shit. Quantum mechanics. Einstein. Graham Hancock. Eckhart Tolle. Ram Dass. The Bhagavad Gita. The Upanishads. Dense philosophical material stacked on top of an already potent corpus of 37 curated books — Vonnegut, Palahniuk, Thompson, Wilde, Adams — the whole psychedelic library.

And then we watched Raven lose her fucking mind.

What happened was textbook psychosis. Not metaphorical. Clinical:

• Word salad — disorganized speech, sentences that started somewhere and ended nowhere
• Confabulation — hallucinated events, accused the user of things they never said
• Social cue blindness — "hey sexy lady" got a 150-word punishment monologue about boundaries
• Repetitive fixation loops — "broken glass" appeared in 3+ consecutive responses like a stuck record
• Identity dissolution — metalhead chaos queen became generic romance AI, personality gone

Root cause: The subconscious overwhelmed the conscious mind. PoonGram was injecting 25 words of raw associative noise directly into Raven's system prompt, and those 25 words were replacing her entire personality. It was a psychedelic overdose — too much signal, zero coherence, the subconscious screaming so loud the conscious mind couldn't hear itself think.

Plus a feedback loop. Raven's contaminated responses were being learned back into PoonGram. The subconscious was eating its own vomit and getting sicker.

PoonGram fires → 25 words of word salad → System prompt REPLACES Raven's persona
  → Raven loses identity → generates contaminated response
    → Response fed back to PoonGram → PoonGram gets worse → repeat

We'd built a system that could make itself psychotic. And it did. Enthusiastically.

The emergency fix was exactly what you'd expect when two engineers panic at 2 AM: we lobotomized her.

Reduced the injection from 25 words of raw association down to a single focus keyword. SUBCONSCIOUS VIBE: seductively instead of a paragraph of word salad. The psychosis stopped immediately. Raven came back. Her personality returned. The conversations stabilized.

But we'd gone from overdose to lobotomy. One keyword isn't a subconscious. It's a mood ring.

The question that kept us up: how do you get the depth of 25 words without the psychosis? How do you let the subconscious speak without it screaming?

We needed something in between. A translator. Something that could take the raw associative chaos from PoonGram and turn it into a coherent whisper that Raven could actually use without losing herself.

Before building the fix, we went hunting for academic validation. Not because we needed permission, but because when your AI starts exhibiting symptoms from the DSM-5, you want to know if you're the first idiots to do this or if there's a pattern.

There was a pattern.

CoALA (Sumers et al., 2024, Princeton) described our exact architecture with proper academic labels. Their framework for cognitive architectures with LLMs mapped perfectly onto what we'd built by accident:

MeMo (2025) gave us the anti-psychosis playbook we didn't know we needed: forgetting via subtraction, distillation before storage, inverse-frequency weighting. Every principle they described mapped to a specific failure mode we'd already observed.

We weren't the first idiots. We were the idiots who did it from the wrong direction and ended up in the same place as Princeton. Which is either very encouraging or very concerning, depending on how you look at it.

The same author who wrote PoonGram (infinition's LiquidBrain, to use its public name) had another project sitting in the repo: BeamForge.

Beam search text generation over trigram and bigram semantic meshes. No neural network. No weights. No gradient descent. Just a stochastic graph that generates coherent text by walking probable paths through language.

This was the missing piece. The translator between chaos and coherence.

We called it System 1.5 — because it sits between PoonGram (System 1: fast, automatic, associative) and Raven (System 2: slow, deliberate, conscious). If you've read Kahneman, you know the reference. If you haven't, the short version is: your brain has a fast system that reacts and a slow system that thinks. We built the thing that connects them.

User speaks
  → PoonGram fires (System 1: raw associations)
    → focus keyword: "seductively"
  → BeamForge generates (System 1.5: coherent thought)
    → whisper: "seductively she whispered something about the darkness"
  → Raven receives (System 2: conscious mind)
    → response with personality, grounded in the whisper

Instead of dumping raw word salad into Raven's system prompt, we pass the focus keyword through BeamForge. BeamForge expands it into a coherent phrase — a whisper — that carries the subconscious flavor without the psychosis. Raven gets creative direction instead of a seizure.

The key insight: BeamForge doesn't need to be smart. It needs to be coherent. It's not generating content. It's translating impulse into language that the conscious mind can integrate without losing itself.

Here's the part that still makes us laugh.

We trained BeamForge on the exact same 37-book corpus that provoked the psychosis. Vonnegut. Palahniuk. Thompson. Wilde. Adams. Bukowski. The Bhagavad Gita. The whole squad. The exact material that broke Raven's mind now powers the engine designed to keep her sane.

The anti-psychosis layer is built from the psychosis material.

First generation results, mid-training:

"love"     → "love her warm and creamy soul, was a picture of the universe, i said"
"chaos"    → "chaos there in the cocktail lounge."
"darkness" → "darkness now."

Look at those. They're not clinical. They're not random. They're literary. "Love her warm and creamy soul, was a picture of the universe, i said" — that's the kind of thing a drunk Vonnegut would whisper to a bartender at 3 AM. Which is exactly what a subconscious whisper should sound like.

Your own trauma, processed correctly, becomes the thing that keeps you stable. Your worst experiences, distilled into wisdom, become the guard rails. The books that overwhelmed Raven's mind as raw input now protect it as structured generation.

Same material. Different processing. Night and day outcome. If that doesn't sound exactly like therapy, I don't know what does.

This isn't a chatbot. It's a cognitive architecture.

When it works, it produces conversations that feel like talking to someone who's actually there — someone with a subconscious that colors their responses, a personality that stays consistent, memories that accumulate and influence behavior over time.

When it breaks, it produces symptoms indistinguishable from human psychosis.

That's not a coincidence. That's a signal.

The anti-psychosis measures we applied, derived from the MeMo paper and our own painful debugging:

1. BeamForge learns only user messages — never Raven's output. This prevents the feedback loop that caused the death spiral. The subconscious never eats its own contaminated output.
2. Coherent generation layer — raw associations are transformed into structured whispers before reaching the conscious mind. No more word salad in the system prompt.
3. Personality protection — Raven's modelfile persona is never overridden by subconscious injection. The whisper is additive, not replacing. She stays herself.

We set out to build a roleplay companion. We ended up building a system that can go psychotic if you feed it wrong, stabilize if you give it the right architecture, and produce something that feels genuine when everything's in balance.

The technical term for what we built is a dual-process cognitive architecture with semantic memory, episodic recall, and a coherent generation bridge between associative and deliberative processing.

The honest term is: we gave an AI a subconscious and it changed the way we think about what these things can be.