Design the memory.
Start with the ability you want, then design a memory structure that can support it.
The architectural principle
Memory can do more
than hold information.
We design how memory works so that storing information also does useful computation. HeatherDB is our first example: memory that can recall from a similar cue.
The idea
We usually think of storage as a place to keep information until a program needs it. Our starting point is different: design the memory itself to do useful work as information is stored.
The structure of that memory determines what it can do. Information lives within that structure; writing it changes how the memory responds. The computation happens in the storing itself.
Storing is computing.
Start with the ability you want, then design a memory structure that can support it.
Writing information changes the memory. The storage operation is already doing computation.
Use the ability built into that memory. In HeatherDB, this means recalling information through associations.
The technical idea is geometry: the structure in which memory lives, designed to support a particular ability. We call the broader approach Storage Native Intelligence.
The architecture
Storage Native Intelligence makes storing information part of the computation. The memory does useful work through the way it is built.
In many familiar AI systems, learned behavior lives in model parameters. Information is stored separately; a model transforms inputs into outputs.
Useful behavior is carried by the model.
Information is stored in a memory designed to do more than hold it. Writing and recalling information are part of how the intelligence works.
The memory participates in the computation.
A familiar detail can bring a whole experience to mind. The next illustration shows how the way memory is stored makes that possible.
See it through HeatherDBOur first example
HeatherDB builds on Pentti Kanerva’s Sparse Distributed Memory (SDM). An experience is remembered across many places. A familiar cue can touch that shared memory and bring an association to mind.
Like remembering something from a familiar detail, recall can begin with a cue that is close rather than exact. The illustration below follows that idea through an everyday experience.
A familiar detail can bring back more than itself.
Remembering an experienceRain → umbrella
Inspired by Pentti Kanerva’s Sparse Distributed Memory , HeatherDB’s foundation. A simplified story of shared memory and recall; the words stand in for encoded patterns.
Imagine walking home in the rain with an umbrella. The experience is remembered across several places in memory, rather than filed in a single box.
One experience, remembered across many places.Later, you hear rain again. The cue is different, but familiar enough to reach some of the same places. Part of the earlier experience becomes active again.
Similar cues reach overlapping parts of memory.Umbrella. The connection comes through the way the experience was stored and recalled. This is the idea Kanerva’s Sparse Distributed Memory helps illustrate: memory can bring something back from a similar cue.
The way memory works makes association possible.Associative intelligence is the first expression.
Storage Native Intelligence is the broader principle.
HeatherDB focuses on associations. Other forms of intelligence need memory designed for different abilities. Storage Native Intelligence is the broader idea: build those abilities into how information is stored.
What this unlocks
The memory and the intelligence it supports can run inside the infrastructure you control. Your information and the ability to use it can stay together.
Your organization's knowledge becomes something you can operate in your own systems.
HeatherDB runs associative intelligence on ordinary CPUs, including a Raspberry Pi 5. For the workloads it supports, you can start with hardware you already own.
That opens a route to useful AI without making expensive GPU infrastructure the first purchase.
Compute requirements shape where you can deploy. Reducing dependence on specialized hardware creates room for different choices: your cloud, a local server, an industrial site, or a device.
Edge intelligence is one consequence. Ownership, local operation, and new ways to build agents and applications are part of the same architectural opportunity.
Where your data belongs
A closer look
The computer still performs work. Writing updates memory; reading combines what has been stored. HeatherDB runs its associative memory operations on CPUs without requiring GPUs. Storing is computing, rather than a step before computation begins.
An index helps locate records. SDM spreads an experience across memory, so a familiar cue can bring it back without an exact match. The illustration shows how that shared memory makes an association possible.
Writing new information updates the memory, changing what it can recall. HeatherDB is designed to accumulate memory without a separate model-training cycle for every update.
HeatherDB is our first example, focused on associative intelligence. Storage Native Intelligence covers the broader idea of building useful abilities into memory. Each further ability needs its own design and evidence.
Go deeper
The foundational research behind memory that can recall from a familiar cue.
The technical foundation behind memory that evolves without training.
Read the original paper and examine the architecture in detail.
How relationships become values a machine can operate on.
Start with your use case
Tell us what you want to solve, the data you work with, and where the intelligence needs to run.