A brain without a soul
That is still the phrase I use for this project.
I am not trying to build another framework that decides how an agent should think, plan, talk, or organize a little company of other agents. I am more interested in the layer underneath all of that: memory, context, ingestion, durable knowledge, and the boring infrastructure that lets an agent avoid waking up as a brilliant amnesiac every time a process ends.
A small model
What the layer is for
An intermediate layer that helps an agent carry its world forward without deciding what its personality should be.
- 01
Ingest
Repositories, feeds, and websites bring scattered knowledge in.
- 02
Contextualise
That knowledge becomes findable around projects and subjects.
- 03
Remember
Durable memory stays separate from one model provider or process.
- 04
Surface
Agents get the pieces they need through their own orchestration.
The arrows describe a direction, not a promise of autonomy.
Why I started thinking about it
I have become increasingly dependent on AI agents for real work, especially coding and research. The models keep improving at a ridiculous pace, but project knowledge is still scattered and context is still expensive. Useful things an agent learned yesterday often have to be reconstructed tomorrow.
Hosted memory products can solve parts of that. Once the memory becomes important rather than convenient, though, cost, control, portability, and trust start mattering a lot more to me.
So I started sketching the foundation I would want underneath whichever agent system I happen to use next year.
It is deliberately less glamorous than an autonomous-agent demo.
What I want it to handle
The project is meant to provide things such as:
- durable ingestion from repositories, feeds, and websites;
- organized knowledge contexts or “banks”;
- long-lived agent memories that are separate from any one model provider;
- queued jobs for slower background accumulation and maintenance;
- ways to preserve what went wrong as well as what went right;
- retrieval and context surfaces that other agents can use without adopting an orchestration framework.
The part I care about most is gradual accumulation. I want an agent to become better informed about a project over time instead of repeatedly receiving an enormous context dump and rediscovering the same world from scratch.
The product questions are not optional
The more intimate and durable agent memory becomes, the less comfortable I am with vague promises about ownership and offboarding.
If somebody trusts a system with years of accumulated context, an account termination cannot simply turn into “good luck with that.” Export, deletion, encryption, backups, billing disputes, policy history, and offboarding are part of the product whether they are exciting to implement or not.
I have spent a surprising amount of time on those questions. I am glad I did.
Current direction
I want to keep the foundation narrow. Provider APIs can sit above it. Agents can sit above that. People can choose whatever orchestration they prefer.
The memory layer should remember useful things and make them available without insisting on becoming the personality using them.
I still like “brain without a soul” as the shorthand.