CognitiveXCOGNITIVEX · MANIFESTO
Your apps have amnesia.
We are building the cure.
Every model you talk to forgets you the instant it replies. We think that is a defect, not a feature. We are building the LCM: software that remembers, learns, and becomes something over time. The memory is the model.
THE PROBLEM
We taught machines to think, and then made them forget.
The large language model is the most capable reasoning primitive ever built, and it has the memory of a goldfish. Every conversation starts from zero. Every preference you stated, every decision you reached, every correction you patiently made, gone the moment the response renders. We have papered over this with longer context windows and retrieval tricks, but the underlying fact has not moved: the model that answers you today is the same frozen function it was a year ago, and it knows nothing about you that you did not just paste into the box.
This is not a small inconvenience. It is the reason AI products feel like brilliant strangers. A colleague who reintroduced themselves every morning would not last a week. Yet that is the experience we have quietly accepted from the software we ask to reason on our behalf. The intelligence is real. The amnesia is the cage.
THE CATEGORY ARGUMENT
The next layer of AI is not a bigger model. It is a remembering one.
The race to scale parameters is real, and it is not the race we are in. Capability is becoming abundant. What stays scarce is a system that accumulates understanding of you and gets better at being yours every day. That is a different category, with a different name.
THE LLM
query → model → response → forget
- Weights frozen at training time
- Each session starts from nothing
- Context is rented, then thrown away
- It answers. It does not become.
THE LCM
query → living memory → reasoning → learning → evolution
- A living substrate that changes with use
- Every session builds on the last
- Memory decays, consolidates, and strengthens
- It learns. Over time, it becomes yours.
We call it the Large Cognition Model. Not because we needed a new acronym, but because the thing we are building is not a language model with a database bolted on. It is a cognitive system in which the memory does the thinking, and the language model is the last, swappable step. The full technical case lives on the LCM page.
WHAT WE BELIEVE
Four convictions we are betting the company on.
- 01
The memory is the model.
The weights are frozen and the same for everyone. The thing that adapts to your life, your work, and your way of thinking is the memory. So the intelligence has to live there, not in the prompt. Encode the structure first. Use a model to render language last.
- 02
The LLM is electricity, not identity.
You plug it in. You do not build your company around it. Swapping one model for another should change output quality, never system behavior. If removing an LLM call breaks the logic and not just the prose, the cognition was never deep enough to begin with.
- 03
Cognition is a loop, not a lookup.
Memory, reflection, pattern detection, decision, learning, and then back to memory. The last stage rewrites the first. Retrieval is one small step inside that loop, not the whole of it. A system that only looks things up will never get to know you.
- 04
Your memory belongs to you.
A system that remembers everything carries a duty most software never had. Living memory is only worth building if it is held with care: yours to inspect, correct, and delete. We treat that as a first-class commitment, detailed on the security page.
THE LONG GAME
A world where software remembers.
Picture the ordinary version of this future. Your tools stop asking the same questions. The assistant that helped you debug last March recalls the fix without being reminded. Your agents share what they learn instead of each rediscovering it alone. The longer you use something, the better it fits, the way a good notebook or a trusted colleague does, until the software you rely on is quietly shaped by the whole history of working with you.
That is the near term, and it is already shipping. Today the LCM offers four-tier memory, pattern detection, salience-scored recall, reflection and introspection, overnight dream consolidation, and cross-agent recall over MCP. Further out, we are reaching toward systems that reason about their own state and grow genuinely new understanding over time. We hold that as vision, stated plainly as vision, not as a claim about what ships today. We would rather under promise the science fiction and over deliver the substrate.
iCog is the first proof: a personal AI with persistent, cross-session memory, built entirely on the LCM. And the same cognition is open to you. The cogx platform exposes it through an SDK, an HTTP API, and MCP, so any app or agent gains living memory in a few lines. Pricing starts free, because the point is to make remembering software the default, not the luxury.
Build software that remembers.
The cure for amnesia is not a bigger prompt. It is cognition that persists. Plug in the LCM and stop forgetting.