Arrisma — Two ways to remember
Recalling a conversation from a year ago

USE MEMORY FAST · WITHOUT CONTEXT ROT

Knowledge from past conversations, brought back two very different ways.

Today's models
TRANSCRIPT 1 YR AGO reading…

Re-read the whole transcript

To recall the conversation, the model reads it again from the top — line by line, token by token. Longer history, longer wait, every single time.

SEQUENTIAL · SLOW · REPEATED EVERY TIME
VS
Arrisma
INSTANT INJECT 1 YR AGO MIND-STATE restored

Inject the saved mind-state

The conversation is kept as a state of mind and injected straight into the model in one shot — the whole memory restored at once, with no re-reading.

DIRECT INJECT · INSTANT · NO CONTEXT ROT
Where the leverage is

Loading one memory is easy. Choosing and merging is the hard part.

Recalling a single past chat is nothing new — plenty of tools do that. Arrisma reaches across all of your past conversations, keeps only the few that matter, and fuses them into one combined state that it injects in a single shot.

01
× 50

Every conversation

All your past sessions, saved as reusable states.

02
5 / 50

Pick the relevant few

Only the handful that bear on the task get selected.

03

Merge into one

Their knowledge is fused into a single combined state.

04
STATE

Inject in one shot

The merged state drops straight into the model, instantly.

The concept

New SOTA for AI

AI model can get updates from the konwledge base every after every ten newly generated tokens, directly into the model's internal state.

1 //
MODEL

Knowledge update goes straight into the model, as often as you want!

RAG loads knowledge directly into the model's state — not retyped as tokens. Words still come in from the keyboard; the knowledge takes a private highway straight into the model's body.

2 //
fixed

Cost stays under control and session costs do not grow with the model state size

A fixed-size context window means the per-step cost never grows during a session — no matter how much content you pour in, the window stays the same.

3 //

Model session can be sent over to anyone in the World

Knowledge and in-progress "AI thoughts" can be packaged and sent over the internet — handed off to any other machine using the same model, which creates an AI ecosystem. Model states can be published for public download from the Internet.

4 //
harness app prompt MODEL

Prompts inject as state too — replacing the harness app

Prompts can be loaded directly into the model's state, just like knowledge. This architecture can stand in for harness apps, giving less tech-agile users comparable functionality without the extra scaffolding.

5 //
older now Jan 1st 20th May Today

Knowledge carries timestamps — a time-aware RAG

Each retrieved fact is tagged with when it entered the knowledge base, so the model knows how fresh every piece is. It can favour the latest information and reason about how things changed over time.

6 //
fixed ∞ one lifetime session

Never-ending session that can run for a lifetime

The model's state carries no growing KV-cache, so nothing piles up as the conversation continues. A single session can stay open indefinitely — one conversation that keeps running for a lifetime, without filling up or slowing down.

WARSAW, POLAND | INFO@ARRISMA.PL