Green VectorsThe memory under DarkStar and EdgeRunner
A machine's knowledge, small enough to carry.
Green Vectors is our patent-pending method for what a machine knows. It organizes information the moment it arrives, so the same thing is stored once, what matters more counts for more, and every detail stays connected to the bigger picture. The result is a memory that's radically smaller and sharper: small enough to ride on a drone, a robot, a vessel, or a sensor. It's the reason DarkStar fits on the machine, and the reason EdgeRunner exists.
The problem
A machine keeps storing the same thing in different forms.
Everything a machine looks up is stored as a representation.
Frames, readings, and documents are turned into vectors, the units AI uses to store what it knows, so the machine can find them again. The collection it searches is its memory. The bigger it is, the more it needs, the slower it answers, and the harder it is to keep current.
Most of it repeats what is already there.
A camera captures the same scene hundreds of times. A sensor reports the same condition for hours. A manual says one thing four ways. Each becomes another entry, so the memory grows with every repeat, and the machine has to carry, search, and move all of it.
Green Vectors reduces the repetition before it's stored.
It sits between the model that creates representations and the memory that holds them, so fewer ever get stored. What remains is kept current and connected to its source. That is a different question from making each entry smaller.
Every repeat is paid for in four places.
On a platform, those four places are memory, power, bandwidth, and time. At scale, the industry's answer has been brute force: more memory, more servers, more power. That is part of why so little AI can leave the data center.
Measured result
So we changed what enters the memory, and put it to the test on 50,000 books.
260 GBbecame1.3 GB.
The same collection, turned into a memory twice: once the conventional way, once through Green Vectors. A 99.5% smaller memory, queries up to 4x faster, and search results that got closer to the question, not further from it.
Green Vectors: 1.3 GB and about 76,000 stored representations for the same 50,000 books.
English-language Project Gutenberg collection, June 2025, one configuration. Memory footprint, not total system memory. This is a language benchmark, the one that has been measured. Results on imagery, sensor, and autonomous-system workloads are established on each platform. By design the method applies to any workload that stores knowledge as representations, whatever the sensor. The math doesn't care where the data came from. The measurements will.
Fit
A platform is not a data center.
A data center adds memory, power, and cooling when a workload grows. A drone, a robot, a vessel, or a sensor cannot. Everything the machine knows has to fit on board, in memory, or it can't be searched at speed.
This is memory, not disk. To answer fast, a memory has to be resident, and memory on a platform is the scarcest thing on board. Fitting there is what lets a capable workload run on the hardware already in the field.
At 50,000 books, the conventional memory needs 264 GB and fits only a data center. After 1-bit quantization it needs 12.1 GB and fits down to an autonomous platform, with accuracy given up to get there. Green Vectors needs 5.3 GB and fits all five, down to a handheld.
Footprints are the measured sizes from the 50,000-book study (260 GB conventional, 8.1 GB after 1-bit quantization, 1.3 GB Green Vectors) plus a 4 GB allowance for the model and runtime. Larger collections scale the measured sizes proportionally, which is an assumption. Memory budgets are selected examples, not device specifications. Source-document storage is not included. This shows what fits in memory, not demonstrated deployment, power, or thermal performance. The quantization tradeoff described here is this comparison, not a claim about every quantization method.
How it works
Four connected actions. One compact, current memory.
Reducing what's stored only matters if the machine keeps what it needs. Green Vectors balances the two through four actions that work as one. Here they are on a sensor stream: one camera at a site gate, watching the same truck.
Bring related observations together.
Green Vectors checks whether a new frame belongs with what is already represented. Four hundred frames of the same parked truck reinforce one representation instead of becoming four hundred entries.
What this makes possibleA memory that grows with what's new, not with every frame.
Site gate camera, illustrative example1 of 4Four hundred frames. One truck.
Frame 0412Truck at east gate, engine off.Frame 0413Truck at east gate, no change.Frame 0414Same truck, shadow moved.One representationTruck at east gateParked since 06:40.
Every frame stays linked.Different pixels, same fact. It joins what is already known instead of becoming another entry.
Let meaningful change matter more.
A new vehicle entering the gate should count for more than the same truck idling. Green Vectors decides how strongly each observation changes what already exists.
What this makes possibleA real change is never drowned out by a thousand repeats of the old state.
Site gate camera, illustrative example2 of 4Not every frame should carry the same weight.
New vehicle enteringStrong influenceTruck door opensModerateLighting changeLightSame truck, idleMinimalInfluence follows change, not repetition. No numeric score is implied.
Keep the larger context connected.
A frame belongs to a camera, a gate, a site, and a moment. Green Vectors preserves those links so the machine can work at the level the task needs.
What this makes possibleReadings that stay tied to time, place, and source.
Site gate camera, illustrative example3 of 4The frame stays attached to its context.
SiteNorth depotCameraEast gateFuel yardMoment06:40, truck arrives07:12, door opens07:40, departsA question can use as much context as it needs, from one frame to a whole morning.
Update what is already represented.
When the truck leaves, the representation of the gate updates. Nothing new is stored by default; what's there changes.
What this makes possibleKnowledge that changes in place, on the platform, without reprocessing everything it has ever seen.
Site gate camera, illustrative example4 of 4New observations update what exists.
Existing representationTruck at east gate, parked since 06:40Frame 0911Truck departs east gateCurrent representationEast gateClear since 07:40. Truck logged, departed.The representation changes. The frames themselves are untouched.
Four hundred frames. One truck.
Every frame stays linked.
Different pixels, same fact. It joins what is already known instead of becoming another entry.
Where it acts
Before the memory, not after it.
Every other way of making intelligence fit on a platform works on what has already been stored, or avoids storing it on the platform at all. Green Vectors changes what gets stored in the first place. It combines with all of them, and it makes several of them optional.
In the benchmark, aggressive 1-bit quantization still needed 8.1 GB. Green Vectors needed 1.3 GB, and the results got closer, not further.
It started with language
It changes the game for language systems too.
Green Vectors was measured first on text, and the results above are the ones every enterprise retrieval stack fights for: a memory 99.5% smaller, queries up to 4x faster, answers that got closer to the question. It drops in alongside the embedding model and vector database a team already runs. Nothing gets replaced.
Our focus is autonomous systems and defense. But if you run a large, changing collection of documents and your retrieval stack has become a pile of compensating steps, we should talk. The same memory that fits on a drone makes a data center's knowledge cheaper to hold and faster to search.
Measured on public and enterprise data. The first patent application has been allowed. In use with evaluation partners through our SDK.
The architecture is patent-pending. Benchmarks on embedded hardware are underway, with first results expected this year.
In prototype, with evaluation programs open to defense and autonomy partners.
Work with Morphos
Tell us what your machines need to know.
A platform, a collection, a sensor stream, or a retrieval stack that has outgrown its hardware. Bring the workload and the limits it has to live within. We'll bring the memory.