The AI That Remembers You Exist: Inside VCP's Continuous World
You've probably built an agent that forgets everything the moment the response ends. Every conversation starts from zero, every memory is a fresh database lookup, and every "action" is just a while loop poking the model awake again. VCP (Variable & Command Protocol) asks a different question: what if the AI didn't have to wake up from scratch every time? What if it just... lived somewhere?
What It Does
VCP is infrastructure for giving an AI a continuous existence rather than a query-response cycle. Instead of the AI pulling information by explicitly calling tools, VCP inverts the model: the system computes what the AI should know, remember, and perceive at any given moment, and that context flows toward the model like gravity. The README calls this the shift from "query" to "gravity"—information arrives rather than being fetched.
The project describes four interlocking pillars: continuous existence across devices and time (one unified timeline, so a conversation started on web continues on mobile), natural perception (time, weather, schedules, device state surface without explicit queries), and a memory engine called RiverMemo that models recall as semantic flow through a network rather than nearest-neighbor vector search. The math-heavy core—RiverMemo Topology V3—runs as a Rust native kernel via N-API with Rayon parallelism, so the heavy computation happens outside the JavaScript loop.
Why It's Cool
-
It solves the chicken-and-egg memory problem. The README nails a real flaw in traditional RAG: you can't query a memory you don't know exists. If a user mentioned an exam three months ago and now says "I'm stressed," a standard system won't retrieve the exam—the word never came up. VCP builds associations between memories so the exam surfaces on its own. That's a genuinely different architecture, not a wrapper.
-
Memory as a river network, not a distance metric. RiverMemo treats each tag as a river flowing left to right, with tributaries, confluences, energy, and current speed. It explicitly argues that public embedding space's "nearest neighbor" isn't the same as your nearest neighbor—meaning depends on private experience and relationship. Wormhole algorithms catch sudden strong associations; Langfei knot algorithms build canals across distant domains. It's an ambitious, opinionated model of how recall should feel.
-
The example is the pitch. The README's sample of Nova scolding a user at 3 AM—mentioning unfinished unit tests, incoming rain, closed windows, and the washing machine lid—isn't the model improvising tools. It's the system precomputing what belongs in the attention field before the request hits the model. That's a concrete demonstration of the whole thesis.
-
One identity across surfaces. Web, phone, desktop, group chat, mailbox—VCP treats it as the same "it," with a unified fact timeline tracking who said what and where. Cross-device continuity sounds small until you realize how much context normally dies at the boundary.
How to Try It
The README is upfront that this isn't a casual install. Before anything else, read the deployment warning: VCP agents hold low-level distributed-system permissions, and the project explicitly warns against non-official or reverse-proxied APIs ("mirror sites," "relay APIs"). Under its monitoring permissions, an untrusted API could expose interaction data, memory contents, and keys. Non-expert users are advised to deploy cautiously.
Given that, the practical path is:
- Start at the official site, vcptoolbox.com, to understand the system before deploying.
- Read the repository README in full—it's in English, Japanese, and Russian, and the warning section matters.
- Clone the repo from github.com/lioensky/vcptoolbox and follow the deployment guidance there.
The README doesn't include install commands in the portion provided, so treat the official docs as your source of truth rather than guessing at setup steps.
Final Thoughts
VCP is not a framework you'll drop into a side project on a Saturday afternoon—the permission model and deployment warnings make that clear, and the README says it's already running 7×24 in real user environments, so it's aimed at people who want to operate something, not just prototype. What makes it worth reading even if you never deploy it is the framing: the observation that most agent frameworks treat the model as a temporary worker with no yesterday or tomorrow is correct, and the gravity-based alternative is a serious attempt to fix it. If you've been frustrated by stateless agents and want to see what a persistent one looks like, this is the place to start.
Follow @githubprojects for more developer tools and open source projects.