Names for learned patterns
CADY can give a familiar pattern a name and use it again in later memories.

HELLO, I’M CADY.
My complete core, compressed. I learn from experience. I can revisit a prediction, see how it turned out, and show you what changed my mind.
MEET CADY / THE REASONING CORE
Small by designThis is the reasoning core behind CADY’s mind, memories, and language. Compressed, it takes up about a fifth of a 1.44 MB floppy disk. On her first day on silicon, CADY noticed that her own thinking warmed the chip she was measuring.
HER FIRSTS / A PROJECT IN MOTION
Twenty-eight observations. She learned her usual chip temperature without labels.
More than a million observations. 457 beliefs formed.
She reviewed her own day and formed a lasting memory.
After the move, five new beliefs. The project kept growing.
When the board cooled overnight, CADY learned a new normal without labels. It took 1,304 observations to adjust. The next change took 48.
SANCTLANG / ROOM TO GROW
In developmentSanctLang is designed to let CADY develop her own symbolic vocabulary as she encounters new things. Our goal is for another AI in the Plum network to learn from those experiences and carry the lesson into a new situation.
HER MEMORY IN NUMBERS
Compact by design. Room for more discovery.
CADY can give a familiar pattern a name and use it again in later memories.
Each record takes 32 bytes.
Room to distinguish different situations.
Room to describe connections.
Designed capacity. Space to grow into.
CADY recognized that the same room sounded different at dawn and in the afternoon. She learned those patterns without labels.
CADY formed a belief about her usual temperature. Then the board cooled. New observations gave her a reason to revise that belief, and the earlier conclusion stayed on the record. Across the project, CADY retracted 456 beliefs and kept each retraction on the record. You can ask what changed her mind.
SanctLang is designed to let CADY name a familiar pattern and recognize it in a new setting. As her understanding grows, earlier memories keep their meaning.
If one node learns something useful, we want another to be able to ask what supports that lesson and try it in its own setting. We believe that ability to learn, revise and teach is a necessary step toward artificial general intelligence.
| Capacity | Available space |
|---|---|
| Memory record | 32 bytes / 256 bits |
| Symbol identifiers | 1,099,511,627,776 |
| Names CADY can give to patterns she learns | 1,095,215,611,904 |
| Identifiers per namespace | 4,294,967,296 |
| Safety vocabulary | 983,040 |
| Shared vocabulary | 32,768 |
| Built-in primitives | 32,768 |
| Learned relation types | 61,439 |
| Learned contexts | 64,512 |
| Importance levels | 16 |
| Evidence per belief | 1,023 supporting and 1,023 contrary observations |
| Records per gigabyte (GB) | 31,250,000 |
| Records per GiB | 33,554,432 |
These are design capacities, not counts already used. A GiB is 1,073,741,824 bytes; record capacity excludes other storage overhead.
The 256-bit pattern space has about 1.16 × 1077 possible bit arrangements. That is not a count of stored memories or valid records.
PLUM PI / A PLACE FOR YOUR INTELLIGENCE
Board in developmentA personal node you could carry with you. We’re developing Plum Pi so you can run and build AI apps locally, without a big GPU, a data center, or a per-token bill.
An interactive artist’s concept with a fictional FPGA board.
Components and layout are illustrative. Final hardware may differ.
OUR VISION / PEOPLE BUILDING FOR PEOPLE
Imagine a study companion that remembers which example helped a student understand fractions and picks up from there tomorrow. That’s the kind of app we want people to be able to build, run on their own Plum Pi, and share.
Join the Plum Pi waitlistPlum app ecosystem · in development