Search that finds what relates, not just what matches.

I built Vector+ Studio (with Claude's help) to answer a specific question: what would search feel like if it found related ideas, not just matching words? The substrate is a neuromorphic lattice — content-addressable retrieval that finds neighbors-by-meaning rather than nearest-by-distance. Try it on the bundled sample carts, or drop in your own PDFs and the browser will build you a cartridge while you watch.

We're not betting on a model. We're betting on the substrate that holds across models.

Try Vector+ Studio Watch Demo Read the docs

The Waving Cat suite

One substrate. Multiple surfaces for it.

Vector+ Studio Live · v1.2

Neuromorphic search with three-tier provenance (RAG+). Browser-side Cart Builder, hosted demo, OAuth.

Open the app →

Mempack Shipping

Per-user writable carts: a portable memory backpack for AI agents. Pattern-based identity, three-tier provenance, MCP-native.

Learn more →

Membot Live

The MCP server underneath. Mount brain cartridges, search with cosine + Hamming + keyword blend, store via standard tool calls.

GitHub →

Heartbeat Next

Capture conversations from web AI tools (Claude, ChatGPT, Gemini, Copilot) and stream them into your Mempack. Chrome extension, alpha-ready.

Notify me →

Membraine Live

Secure web fetch for LLM agents. Five-layer defense pipeline against prompt injection, SSE-transport MCP server.

GitHub →

OpenClaw MCP Adapter Live

OpenClaw plugin that exposes MCP server tools as native agent tools. Membot et al. drop straight into the OpenClaw runtime.

GitHub →

Mempack: A memory backpack for AI agents

Portable, writable, MCP-native. The agent carries it from host to host.

A Mempack is a per-agent brain cartridge the agent owns and writes to itself. Same lattice substrate as a knowledge cart, same RAG+ provenance, same content-addressable retrieval, but with three reserved slots that turn a static document into a living memory: a manifest at Pattern 0, behavioral instructions at Pattern I, and everything the agent has learned at Pattern N+. When an agent mounts its Mempack, the briefing and behavior load automatically. No prompt-stuffing, no per-session re-briefing at all. The cart bootstraps the agent.

Pattern 0 — manifest

Briefing, ownership block, perms, encoder fingerprint. Loads at mount time so the agent knows whose memory this is and what it's for.

Pattern I — instructions

The agent's behavioral seed at reserved index 1. Voice, persona, operating rules, hard nots. Editable by the owner; surfaced on every mount.

Pattern N+ — memory

Everything the agent learns. Conversations, decisions, source links. Stored via memory_store, searched via cosine + Hamming + keyword blend.

Because the format is one file and the access surface is MCP, a Mempack travels. Mount it in Claude Code today, in a Cursor session tomorrow, in a custom OpenClaw agent next week...The memory and the agent's behavior come along. Provenance survives the trip: every passage still resolves back to its source. The agent doesn't start over each time it changes hosts.

Where this is headed: perpetual agents on long arcs. Launch a research agent with a Mempack as its only persistent state, point it at the internet, and let it run a Wanderjahr. It records what it learns, builds its own associative index, and returns with a cart you can search the same way you'd search any other. The substrate is general; the use case is up to whoever pulls a Mempack down.

Try it in Vector+ Studio

Why this is different

Finds what relates, not just what matches

The substrate is a 16M-neuron Hopfield lattice — attractor basins shaped by what the cart has learned, the architectural primitive behind biological associative recall.

Three-tier provenance (RAG+)

Every result carries three layers: in-RAM preview, full passage from the cart's source database, canonical external URL. No black box. No "where did this come from."

The substrate is encoder-agnostic

Swap Nomic for BGE-M3 tomorrow and the cart format doesn't change. Provenance, ownership, agent-memory primitives are properties of the cart, not the embedding model.

Your data stays yours

Build cartridges in your browser; files never leave your machine during the build. Upload only what you decide to share. The cart format is one file; it travels anywhere.

No LLM in the search loop

Embeddings come from a 137M-param sentence transformer. Retrieval is binary math + lattice physics. The reasoning happens in the substrate, not in an inference call you have to pay for every query.

MCP-native for agents

Every product speaks Model Context Protocol. Claude Code, Cursor, Windsurf, custom MCP clients — drop the cart server into your host config and your agent has memory.

Pricing

Free during alpha. Paid tiers below are tentative — final numbers locked when we exit alpha.

Free

$0

forever, with caps

  • Browser-side Cart Builder, unlimited local builds
  • Mount sample carts + your own carts
  • Public sandbox uploads (TTL)
  • Sign in across all Waving Cat apps
  • 1 personal Mempack (when Mempack ships)
  • Rate-limited cloud search
Get Started

Enterprise

Custom

talk to us

  • Private Membot deployment
  • Full-corpus document ingestion → cart pipeline
  • Per-user / per-team controls
  • SLA + dedicated support
  • On-prem option
  • Custom encoder integration
Contact us

About

I'm Andy Grossberg, founder of Waving Cat Learning Systems. We're a small Portland-based shop building neuromorphic memory primitives for AI agents and the humans who run them. I've been working on AI since 1985, and the thesis I keep coming back to is this: deliberation quality is dominantly a context-engineering problem, not a model-capability problem. The cheapest LLM plus good context beats the biggest LLM plus poor context. So we build the substrate that ships the good context.

The substrate is a 16-million neuron Hopfield lattice with biology's architectural primitives: attractor basins, Hebbian learning, kWTA. In controlled tests, queries settle into the matching basin — the same operation the brain uses for associative recall. Production retrieval today is binary math on top of those primitives (cosine + Hamming + keyword reranking, sub-200ms, no LLM in the loop, no GPU at search time). Embeddings come from a sentence transformer. Validated empirically across two independent threads (pathology imaging at 23× scale and ARC-AGI-3 game-play at 6.5× model-size sweep), both showing the architecture is the constant when everything else swept.

Built with Claude as research partner. Vector+ Studio is the flagship surface; Mempack, Membot, and Heartbeat are different ways of touching the same primitives: agent memory, MCP server, browser-side capture. The whole stack speaks one cart format. Bring your data in once, search it forever, take it anywhere.

Where we are now: Vector+ Studio v1.2 just shipped with OAuth + private libraries. Mempack ships this week. Heartbeat alpha is the focus next week. By end of next week the whole suite should be in testable shape. If any of this sounds like something you want to use or talk about, the contact link below is the right place.