AGPLv3 open coreSingle-tenant per userHosted in GermanyExport anytime
why afair exists

The tools change. Your context shouldn't. Here is what afair does about it.

What afair is, how it's built, what it draws from. One memory for your work, the people in your life, and everything you'd rather not re-explain, across every AI you use.

what afair is

One machine. One user. Yours.

It runs as a single dedicated machine, one user only. Self-host it yourself, or let afair.ai run that machine for you in the EU. Either way it is single-tenant by design: your data never shares a database with anyone else. It connects to your AI tools through MCP, the open protocol every major model now speaks. Any client that talks MCP can read and write to it.

It stores text, PDFs with their text layer extracted, audio transcribed with Whisper, screenshots and photos described by vision models. A work contract, a long message from a friend, a voice note about your kid, a photo of a whiteboard, all of it. Up to one gigabyte per file. Everything is retrievable by its contents, not just its filename. You can export the whole vault at any time.

host
EU
tenancy
one user, one machine
protocol
MCP v1
per file
up to 1 GB
export
full vault, any time
the architecture

Three layers, each mirroring a brain memory system.

The brain mapping is real, not a metaphor bent for marketing. Every memory product already out there got a close look, and a few of their ideas carried over. But most of what a memory actually needs was missing, so the design went where the problem was already solved: the brain. Each layer maps to a memory system the brain runs.

i
MCP layer
≈ thalamus

Gates and relays everything that comes in and goes out. Three verbs in v1: remember, recall, observe. The thalamus does the same for sensory input, deciding what reaches the cortex and routing it there.

rememberrecallobserve
ii
Intelligence layer
≈ triple network

Inspired by Menon's salience, central-executive and default-mode triple. A salience worker scores every event for what matters. A mode-switcher routes between focused CEN for immediate attention and wandering DMN for integration.

iii
Storage layer
≈ hippocampus + cortex

An append-only event log, content-addressed, never overwritten, with an interpretation layer on top: extracted facts, entity graphs, embeddings. It mirrors Complementary Learning Systems, the fast sparse store and the slow consolidated one.

what it draws from

The frameworks.

Three bodies of work shaped the architecture. If you know them, you will recognise the seams. If you don't, here is the short version.

Complementary Learning Systems
McClelland & O'Reilly, 1995

The brain keeps two memory systems, not one. A fast hippocampal store writes single experiences down sparsely and quickly. A slow neocortical store integrates them over time into stable structure. afair separates the same way, so new input never has to overwrite what is already settled.

Predictive Processing
Friston

A brain is a prediction engine that updates on the gap between what it expected and what arrived. Surprise is the signal worth storing. The salience worker scores events on roughly this principle: what was unexpected, and what does it change.

Society of Mind
Minsky, 1986

Intelligence is not one process. It is many small agents, each simple, none in charge, producing mind through their interaction. The intelligence layer is built as workers in that spirit: small, specialised, coordinating rather than commanding.

how it stays useful

The interface is frozen. The brain underneath isn't.

Three verbs. remember, recall, observe. Defined on day one and not going to change. Your AI tools are written against those three, so they keep working.

Underneath, the substrate keeps improving. Sharper extractors. Better salience. More accurate entity resolution. A surprise score that learns what to elevate. New workers come online when older ones are outgrown. afair watches its own operation and adjusts.

Every change earns its way in before it ships, run against your own vault first. Three independent models from Anthropic, OpenAI, and Google compare the variant to current behavior on real past events. Today the loop runs in observation mode while the grounding in concrete recall metrics is built out; when it lands, only majority wins go live. A separate monitor watches the first events after each change. If signal degrades, the change reverts. Recorded. Reversible.

All of it stays out of your way. The same three calls keep returning better answers over time, with no migration and no version to pin yourself to. The brain learns, the contract holds.

what makes it different

Vendor memory remembers for the vendor.

vendor memoryafair
who owns the datathe vendoryou
works across AI toolsno, only theirsyes, any MCP client (including the next one that hasn't been built yet)
tenancyshared, multi-tenantone user, one machine
schemafixed by thememerges from what you put in
can you see and correct itopaque; weight-baked memory, not at allyes, it's your data to read and fix
data residencywherever they hostEU by default
right to exportpartial, opaquethe whole vault, any time

The major labs face a structural pressure here. Portable memory weakens lock-in, and lock-in is part of how they hold a user. Anthropic created MCP for exactly this kind of interoperability.

A newer bet goes further still: train your context into a model's weights so it just knows you. That buys speed and the deepest lock-in there is, because memory in weights cannot be read, moved, or corrected. The more the field pushes that way, the more a memory you can open, carry, and fix is worth holding onto.

afair sits beneath MCP rather than against it. It is the substrate an MCP client connects to when it wants memory that belongs to the person, not the platform.

further reading

Further reading.

  1. McClelland, McNaughton & O'Reilly (1995). Why there are complementary learning systems in the hippocampus and neocortex. Psychological Review.
  2. Kumaran, Hassabis & McClelland (2016). What learning systems do intelligent agents need? Trends in Cognitive Sciences.
  3. Menon, V. (2011). Large-scale brain networks and psychopathology: a unifying triple network model. Trends in Cognitive Sciences.
  4. Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.
  5. Minsky, M. (1986). Society of Mind. Simon & Schuster.
hosted, coming soon

Own the memory behind every AI you use.

Leave your email for the first invite, the day hosted afair opens. One message, nothing else. Or self-host the open-source core today.

By submitting, you agree to receive a confirmation email and a single welcome message. Your email address is processed as described in our privacy notice and you agree to our terms.