The Problem
Every AI memory architecture trades against your inference budget.
LLM-mediated memory
Mem0 · Letta · Zep · Graphiti
Mem0 reports 7–8 s per memory recall; Zep ~4 s. Every operation costs that latency and the LLM tokens consumed by extraction, summarization, and routing — on top of any tokens the consuming agent injects downstream.1
Standard vector-DB RAG
Pinecone · Qdrant · Milvus · pgvector
Avoids the memory-layer LLM cost, but production midsize implementations still inject 8,000–12,000 tokens of retrieved context per query. At 100K queries/month: $3K–$5K on Standard tier; $30K–$180K on Premium.2
Long context windows
1M-token models
Inflate inference cost, lose information past ~16K tokens (lost-in-the-middle3), and vanish at session end.
All three share a structural problem: memory is a database lookup paid for at inference time. There is no architectural escape within the database paradigm.
1Vectorize.io 2026, Mem0 vs Letta (MemGPT); DEV Community 2026, 5 AI Agent Memory Systems Compared.
2Redis 2026, LLM Token Optimization; Towards AI 2026, LLM API Token Caching; current Anthropic + OpenAI public per-million-token pricing.
3Hugging Face 2025, Advanced RAG cookbook.
The Solution
A portable, 1-bit memory substrate. Production-validated retrieval verified across three independent fleets and one clinical dataset.
Products
Surfaces
Vector+ Studio
Membot
Membraine
Mempack
MITOpen source
F1
Retrieval
+Cosine
⊕Sign-bit Hamming
★Keyword reranking
F0
Memory
Hebbian
kWTA
Attractor primitives
Encoding research active
Substrate
Lattice
128 MBfixed
1-bitneurons
1-bitweights
XOR + popcount
Substrate doesn't grow when storing more patterns
Hebbian primitives during imprint. No backpropagation, no gradient descent. Cart-stacking architecture sidesteps catastrophic forgetting by design.
Hippocampus episodic linking (PREV/NEXT) woven into the same substrate for sequential navigation alongside semantic retrieval.
Open source MIT (Membot, Membraine). Substrate is free; products built on it monetize.
Encoder-agnostic. Anything you can build an encoder for — text, image, audio, video — drops onto the same lattice.
The Numbers
Validated. No degradation visible
Two live demos. at the tested upper bound.
Live · Membot
Semantic search · 2.4M arxiv passages
project-you.app/membot/app
→
Live · Vector+ Studio
Drag-drop cart builder · RAG+ provenance
project-you.app/vps/app
→
One continuous run up to 3 million patterns.
And we still haven't found the ceiling...
And we still haven't found the ceiling...
| Metric | Result |
|---|---|
| Patterns stored, single lattice | 3,000,000 · 9-checkpoint validated · no upper bound found |
| Recall fidelity (R@1) | 1.000 at every checkpoint, 1K → 3M |
| Tokens per memory op | 0 LLM tokens · sub-200ms retrieval |
| Search speed at 3M patterns | 2.1ms (BALANCED) · constant with scale |
| LongMemEval R@5 | 95.0% · no LLM in retrieval pipeline |
| ARC-AGI-3 | 84.9% · 23/25 games · ~$250 total cost |
| Sign-bit arithmetic empirical floor | 24,926× signal vs random · 370K vocaba |
| Histopathology pilot (multimodal) | 100% tumor sensitivity + 100% release-tier specificity at 93,522 patches, 4 encoder consensus |
| BEIR retrieval | Sign-zero blend matches cosine within 1pp at 32× compression |
| Production deployment | $28/mo droplet · multi-agent serving |
04-28-26 · sign-bit arithmetic empirical floor, 370K-vocab. Encoders validated in pathology pilot: CONCH, UNI, Virchow2, Phikon-v2.
Market
Vector retrieval is our wedge. The substrate's reach is broader.
Vector database market · 2025 → 2030 · CAGR 27.5%1
Standard sign-zero retrieval at 32× compression matches Pinecone, Weaviate, Qdrant, Milvus, Zilliz — with no reindex-on-write and zero LLM tokens in the memory operation.
Adjacent · Cloud DBaaS
$57.5B by 2028
CAGR 22%2 · Vector DB sits inside this. Slice of a fast-growing slice.
Adjacent · Graph DB
$2.14B by 2030
FK-as-edges + episodic hippocampus map onto the same substrate.1
The substrate is encoder-agnostic. What we ship is shaped like a vector DB; what we built is a binary-pattern memory layer that takes anything you can build an encoder for. Each new encoder unlocks an adjacent market.
Text
Production
Nomic embeddings · live at project-you.app
Images
Pilot
4 vision encoders · 100% tumor sensitivity
Audio
Demo
Mel spectrograms · paired-pattern recall
Video
Demo
Cross-modal cue → modality recall
1MarketsandMarkets, Vector Database Market — Global Forecast to 2030 (12-25); Graph Database Market (01-25).
2MarketsandMarkets, Cloud Database & DBaaS Market.
Savings
No LLM in the memory layer. Watch what that does to your bill at scale.
Mem0/Letta/Zep-style memory systems pay LLM tokens twice per query: once for the memory operation itself (extraction, summarization, routing), and again when retrieved content is injected into the consuming model's prompt. We eliminate the first cost entirely.
| Usage volume | Efficient (Haiku / Mini) | Standard (Sonnet / GPT-4o) | Premium (Opus / Pro) |
|---|---|---|---|
| 1 query | $0.002 – $0.05 | $0.03 – $0.22 | $0.30 – $1.80 |
| 1,000 queries | $2 – $50 | $30 – $220 | $300 – $1,800 |
| 100,000 queries / mo | $200 – $5,000 | $3,000 – $22,000 | $30,000 – $180,000 |
Savings range applies most directly to teams using LLM-mediated memory (Mem0, Letta). For teams using standard vector-DB RAG, savings come from higher-precision retrieval allowing fewer chunks per query — typically 3–5× reduction at the lower end. Cost math compounds either way; magnitude depends on what you're switching from.
Savings model: 10,000 input tokens eliminated per query · Anthropic + OpenAI public per-million-token pricing as of 04-2026.
One Substrate, Three Surfaces
Same substrate, three monetization surfaces.
Primary
Developer / DBMS surface
Vector+ StudioHybrid DBMS · cart-as-primary-object
- Free hybrid DBMS — semantic-first, SQL as power tool
- Cart Marketplace as agent-commerce primitive · ownership-key auth doubles as DRM
- Drag-drop document ingestion → branded cartridge → portable artifact
- Tier model: free hosted demo · per-cart · hosted enterprise
Enterprise infrastructure
MembotSelf-hostable substrate
- Single-user · federated · multi-user (Membox) modes
- The first neuromorphic DBMS built around architectural primitives (lattice, sparse coding, Hebbian), to our knowledge
- Customer-support KB · internal company database · agentic memory layer
- Tiered pricing per deployment
Consumer
HeartbeatChrome extension · 7 LLM platforms
- Captures chats from ChatGPT, Claude, Gemini, Mistral, Grok, Copilot, Heart
- Cross-device sync via Supabase
- "Your LLM chats stop being goldfish-brained"
- Free tier + paid sync tiers
Vector+ Studio
free hosted demo / per-cart / hosted enterprise
Membot
self-hosted free / hosted enterprise
Heartbeat
free tier / ~$5/mo sync
Three surfaces, three revenue motions: marketplace take-rate · enterprise-hosted tiers · consumer freemium. One substrate, one cart format underneath.
Hardware pathBuried lede — see slide 09
Binary operations (1-bit neurons, 1-bit weights, XOR + popcount) map directly to FPGA → ASIC silicon. Hardware partnerships in discussion. A memory chip, not a software library.
FPGA → ASIC
Royalty + OEM licensing
Royalty + OEM licensing
Go-to-Market
Open source substrate is the distribution. Monetization layers sit on top.
01Running today
Open source flywheel
Live now- MIT-licensed substrate at github.com/project-you-apps (Membot, Membraine, Looper)
- Two live public demos: Membot (2.4M arxiv) + Vector+ Studio
- Sign-bit-arithmetic + pathology results posted to LinkedIn
ComparableHugging Face open-model hub flywheel
02Marketplace
Cart Marketplace
V1 → V1.2 · Q3 2026- Sellers package carts: USPTO patents, SEC EDGAR filings, public health datasets, vertical KBs
- Buyers download with ownership-key auth (DRM via authentication, not encryption)
- Take rate on transactions; recurring revenue from cart-as-product
- Network effects: more carts → more buyers → more cart-makers
ComparableHugging Face model hub + Zapier integration marketplace
03Enterprise
Hosted enterprise
V1.2 · Q4 2026 Validated demand · 04-24-26- Self-hosted Membot is free (MIT). Hosted managed Membot is paid.
- Tiered: per-cart-month · per-query · per-seat
- First external pitch: consulting shop with mid-market book asked to deploy both internally AND for a specific client engagement before the consumer demo finished. Technical deep-dive scheduled.
ComparableGitLab — free self-hosted + paid SaaS
04Consumer · parallel
Heartbeat consumer
Q4 2026- Browser extension free tier — capture from 7 LLM platforms, local search
- Paid tier: cross-device sync via Supabase, semantic search across every LLM chat ever had
- Freemium · ~$5/mo for sync
ComparableNotion / Obsidian Sync
Hardware (long horizon, V2+): FPGA → ASIC partnership. Royalty model on silicon, OEM licensing for embedded substrate (smart home, automotive, healthcare devices). Long-term floor — not the immediate revenue path.
Roadmap & Hardware
V1 today (80% shipped). V1.2 next quarter. V2 by year-end. Silicon as the long horizon.
V1Q3-2026
80% complete · live at project-you.app/vps/app
- Vector+ Studio Hybrid DBMS surface: drag-drop cart builder, RAG+ provenance, sandboxed upload-your-own-cart, three-layer cart-format RWX
- Cart Builder V2 Browser-side WebGPU build pipeline. Substrate smoke-tested 05-06-26 · server-parity at 6th-decimal drift. Your data never leaves your machine.
- SQL editor v1 Toggle alongside semantic search · ~10 commands wired to existing Membot API
- OAuth Supabase Auth + Google/GitHub/Apple · platform-wide identity
- Cart Marketplace beta Sellers list · buyers download with ownership-key auth
V1.2Q4-2026
Cognition + first enterprise customer
- Cognition Engine substrate Free-wake, daydream, predictive coding, concept-cluster Hebbian (CCs filed)
- First paying enterprise customer Pipeline from 04-24-26 consulting team · deep-dive scheduled
- Pathology consortium expansion Phase 1: 100% sensitivity in hand · Phase 2 adds NCT-CRC + PANDA stratified, cross-tissue
- Lattice-LM experiments Bigram/trigram prediction floor · Spec 06 pre-registered
V2
Marketplace at scale + silicon prototype
- Cart Marketplace at scale Payment rails · take-rate revenue · agent-commerce primitive
- SQL ingestion .sql files as cart source. Rows-as-passages, schema-as-manifest, FK-as-hippo-edges
- Cloud storage backend Cloudflare R2 default · pluggable CartStorage interface (S3, GCS, user-hosted opt-in)
- Silicon prototype FPGA build begins (partner TBD)
3Precedence Research, Edge AI Market Size to Attain USD 165.05 Billion by 2035 (current).
The Ask
$25–50K bridges us to the pre-seed round.
Founder · this is a founder-bet at this stage
40 years of personal iteration on this architecture. Original concept hit paper in 1985. Production runs today on a $28/month droplet, serving multi-agent traffic with 2.4M passages at sub-200ms.
Solo technical founder + Claude as co-developer. Early instance of an LLM listed intentionally as co-author on an academic submission (ARC-SAGE: Palatov, Grossberg, et al., in preparation). Claude Code uses LatticeRunner's substrate as its own memory layer — eating our own dogfood at the daily-development level.
What $50K specifically buys · 6-month bridge
| Allocation | Amount | Deliverable |
|---|---|---|
| Andy runway | $25K | 6 months focused build time at bootstrapped pace |
| Hardware engineer (contractor) | $10K | 2 months: GPU optimizations on the lattice runtime + breadboard XOR rough demo (64×64 L2 or 256×256 L3 scale). Physical silicon-shaped artifact in hand — turns "interesting concept" into "show me the silicon, here's the silicon." |
| Hardware partnership talks | $5K | Travel + prototype hardware deposits |
| Hosting + marketing | $5K | V1 launch (Reddit / HN / Product Hunt) + deeper droplet capacity |
| Legal | $5K | Contracts, provisional patents, etc. |
End-of-bridge milestones · the angel's "I bridged them to the real round" moment
M1V1 shipped
with Cart Marketplace beta live
with Cart Marketplace beta live
M2First paying enterprise customer landed (pipeline from 04-24-26)
M3Cognition Engine V1.2 prototypes running (CCs filed; OAuth + cron foundation)
M4Heartbeat early growth — ~1K users
M5Breadboard XOR demo of the lattice in hand — physical-substrate artifact ready for hardware-partner conversation
Try it right now
Pre-seed round bridging to: $500K–$1M · 12–18 mo runway → V2 + Series A
Andy Grossberg · andy.grossberg@gmail.com · wavingcat.dev · github.com/project-you-apps · Discord: andygrossberg