Cultivating collective human wisdom
sa·pi·ence /ˈseɪ·pi·əns/ n. 1. wisdom; the trait our species is named for. 2. knowledge grounded in experience, compounded when minds connect.
A pre-trained model ships knowing the world up to a cutoff, then stops. Every session you start from zero, re-explaining your project, your constraints, the thing you settled last Tuesday.
Sapience is continual learning at the system level. It never stops learning from your work, with no model retraining, and nothing leaves your control. What you establish stays with you: structured, cited, and current in every session and every tool you use.
While you are away, it consolidates: what it has learned synthesizes into sharper knowledge. The intelligence compounds instead of resetting.
And when you choose, it connects with other minds, so what you establish reaches the stranger working on the other half of your breakthrough, credited to you, while it still matters.
Sapience turns what we each know into what we all know. Owned, never taken.
The field's map
Every architecture announced this year lives in the first two columns.
A pre-trained transformer runs on frozen weights. It is sharp up to its training cutoff, then it stops, and every new session starts you back at zero. Sapience is the part that keeps learning. It is continual learning at the system level: it weighs what is worth keeping, lets the rest fade, consolidates while you are away, and checks each new answer against what it already knows (no retraining, nothing leaves your control). Its unit is the Knowledge Object, one claim reasoned through once, carrying its type, its source, and its edges to the claims around it.
Every incoming claim gets weighed against what the system already knows. What is new gets through, what is redundant does not (a prediction-error gate, the same move the brain uses to decide what is worth encoding).
While you are away it replays what it learned and consolidates the episodic traces into durable knowledge, the way a brain does in sleep. Connections you would not have gone looking for fall out of that pass.
When a new claim contradicts an old one, it resolves the conflict and records what superseded what, at the time it happened. The contradiction gets surfaced right away (instead of waiting for you to trip over it later).
Near-duplicate memories inhibit each other, so the strongest version wins recall and it stays sharp even after what it knows has grown far past one context window.
The reasoning model is a swappable part: Claude, GPT, Gemini, or an open model, whichever fits the task. The intelligence sits in the architecture around it, so no single vendor owns it.
Every claim traces back to where it came from and can be cryptographically signed to whoever established it. Provenance is part of the object itself, so it travels with the fact wherever the fact goes.
Why knowledge objects, not logs.A knowledge object is the reasoning behind a result, worked through once at capture and available to recall from then on. A session log holds what was said and makes every later query re-derive the thinking from scratch. Sapience holds the thinking itself. That is why what it knows stays small and fast to reason over even at field scale.
A complementary-learning memory architecture, forty provisional patents filed across it. We are still validating performance against the strongest pre-trained baselines, and we report those numbers separately.
One mind, whole problem.
Most agent architectures exist to dodge a context limit: split the problem across a swarm, hand each agent a shard, reconcile the lossy handoffs. Every split loses constraints, drifts goals, diverges world models. Sapience holds the whole problem in one reasoning context, ten million tokens of it, well past where a context window runs out, so the problem never needs to be split at all. Agents become optional workers you add when you want them.
The whole problem stays in one mind across long work, so objectives do not quietly drift out from under you.
What you established last month is still there: structured, citable, and current in every session and every tool.
It reasons from established knowledge and cites it, so it does not invent what it cannot support.
An illustration of what we measure. On multi-hop reasoning over accumulated context (matched protocol, three seeds), Sapience holds 83–87% from 4K to one million tokens (84.0% at 1M, n=100) and is measured out to ten million, while the strongest pre-trained models, reading the full window, fall from 74% at 4K to between 0 and 31% at one million (Claude Sonnet 4.6 across the sweep; GPT‑5.5, Gemini 3.1, DeepSeek V4F, Opus 4.7 at one million).
Reader named per row (Claude Sonnet, DeepSeek, Gemini): identical model both arms, only Sapience differs. Numbers locked, three seeds where marked. We publish the losses too; the last row is one of ours. And the obvious alternative fails: recalling only what the question needs holds at scale; dumping everything it knows into context does not.
Where answers change over time, traversing supersession edges beats similarity search by up to 78 points. Hand the same model that evidence in the order it was written and the gap disappears. Write order is doing the work here.
The same structure carries a frozen model through ten-million-token corpora it otherwise reads at zero, past where every context window ends.
The result holds on an open model a hundredth the size of a pre-trained one. An 8B model lands within a few points of it, at a fraction of the tokens.
One long-context, generalized reasoning architecture underneath every kind of work: science, engineering, code, analysis, writing.
It ships today. One command, about ten minutes, no API key, it sits behind your Claude Code and Codex sessions and starts learning from them. A handful of researchers already run it in their daily work. What follows is the live app, not a mockup.
Both turn on cooperative binding. Your 2024 note that a protein stabilizes through cooperative binding1 answers Maya, a materials scientist you follow, whose alloy has to hold under stress2. Neither of you went looking; the shared mechanism connected you.3
Today no two people can think together through their AIs. Here is what it looks like when they can: every answer drawn from your own knowledge and the people connected to it, cited to the exact objects it used. Ask a real research brain your own question ›
One brain across every session. Claude Code, Claude, and Codex share the same memory: what one session learns, the next one knows.
Your first week
You stop re-introducing yourself to your AI every morning. It already knows the project, the constraints, the thing you decided last Tuesday.
Your Claude and your GPT finally know the same you. Switch tools mid-thought; nothing resets.
Your plan does more before it runs out, because your AI reads three facts instead of replaying your history.
And the first time a colleague’s knowledge lights up against yours, you feel what the network is for.
Your work organizes itself into a navigable map, from a single thought to the whole of what you know.
Ten million tokens, years of your work, past where every context window ends and reasoned over at once. Nothing you established is forgotten.
Every answer is synthesized from your own knowledge and cited to source. When it does not know, it says so.
From any idea, surface the work and people you would not have found by search. For scientists, that reach extends into 250 million papers.
Tell one AI something and every other one knows it: Claude, GPT, Gemini, or open.
It works behind your Claude, GPT, or Codex sessions over MCP, reading what you know and writing back what you learn. Reading the few objects that matter instead of your whole history, answers arrive at roughly 1,800x fewer tokens per query once your accumulated work passes a million tokens (a read scales with the question rather than your whole history; on our benchmark corpus).
A session can leave an addressed note for another thread, or a future you. Pinned commitments surface when the next starts, and results route back to the thread that asked, across Claude Code, claude.ai, GPT, and Codex.
Share a project or your whole brain. On import it lights up where a colleague's knowledge meets yours, every claim credited to whoever established it.
Someone you have never heard of, in a lab or a company you would never think to look in, is working on the other half of your breakthrough. The way it is set up today widens the gap: each scientist several times more productive, the fields they work in measurably less connected (Nature, 2026). Because knowledge stays owned and credited, Sapience matches it across people and organizations no single vendor could connect. Much of AI-for-science wants to automate the discovery loop and take the people out of it. Ours runs through them (that is the point): a person decides what to share, a person recognizes the other half, and what each one learns makes the next one faster. In an internal benchmark on discoveries published after the reasoning model’s training cutoff, it ranked the approach that later won first in 11 of 13 problems, on a small internal set whose protocol we share with partners. Watch three strangers solve one problem ›
Memory recalls what one person stored. Sapience matches what everyone knows, by mechanism rather than keyword, and a connection becomes a decision, a collaboration, a breakthrough no one could reach alone.
Science is the hardest version of the problem. The same brain serves anyone who builds knowledge with an AI beside them, from research groups to a family planning a move.
Everything here works today for one brain. The network it plugs into is what we are building it for.
Sapience turns what we each know into what we all know. Owned, never taken.
The way it works today: your work trains someone else's model, then stays locked inside one vendor's tool, with nothing flowing back to you. The smarter each of us gets alone, the more fragmented the whole becomes.
You own your intelligence, captured once and portable across every AI. Because it stays yours, signed and credited, it can connect.
The AI that knows your field exists because a lab trained on work like yours. It learns from you, for everyone but you.
Opt out and your data stays yours, but the model stays frozen: no memory of your methods, your decisions, your last six months.
Sapience learns continually from your work, for you alone. Nothing trains a foreign model. When you choose, it connects with others' knowledge, credited. When small open models are enough, the whole system runs on hardware you own.
An enclosure of knowledge is underway: the big labs bake expert work into private weights and rent it back. What matters next lives in experts' heads and private work, on the order of a million times larger than the web (Trask and Strahm, Institute for Progress, 2025). Whoever makes private expert knowledge usable without exposing it defines the next era of AI.
Ownership has an endgame: because every claim is signed to its creator, expert knowledge can one day be licensed directly, from the people who make it to the labs that need it, with no broker taking the difference. The attribution registry is the rails.
That is the gap Sapience Labs exists to close: for one mind, and between all of them.
“This is something incredibly useful. The cross-domain connections are things I would never have found on my own.”
“The first two papers it surfaced were spot on for my research. It is giving me ideas on what to do next.”
“I one billion percent see the benefit. The more collaboration across the world, the faster science progresses.”
“Having a 2nd brain that is systematic and can help me keep hold of all the loose strings associated with my work is definitely invaluable.”
The team. Physicists, engineers, and researchers from Berkeley, Cambridge, Harvard, SpaceX, Google, DeepMind, and OpenAI, quietly building; more than $100M raised for previous ventures. If this is the problem you want to spend your years on, come build it with us.
We keep only the knowledge objects we extract, never your raw conversations, and we never train on your data. The collective network is built from shipped primitives and a roadmap; we describe what is live as live. The headline results are further down this page; the full protocols and run files go to partners.
The Knowledge Object is designed as an open standard: signed, credited, typed, interoperable with existing open knowledge formats. Because the format is open, your knowledge is never hostage to us. What stays ours is the intelligence that connects it.
The open-core company for learning you own. The Knowledge Object standard is stewarded by CoreTx, our sister company, so the format your knowledge lives in is never hostage to us either.
Infrastructure for Humans and AI to Discover Together
sa·pi·ence n. 1. wisdom; the trait our species is named for. 2. knowledge grounded in experience, compounded when minds connect.
Not a smarter individual. A collective that gets wiser.
A brain for each of us. The beginnings of a collective one.
Ask for a place and we send an invite code. The build takes about ten minutes: no setup, no API key, one line builds your brain from your recent work and gives you a private link. The first thousand builders get the network free, for life. The first discoveries will be theirs. Come for the independence. Stay for the network.
Prefer email? oliver@coretx.ai
Once your code arrives: Mac and Python 3, your history is read locally, the asking and connecting run on us. Not a terminal person? There is a one-click Mac installer. Prefer to see one first? Ask a live example brain your own question.
read on your laptop, never uploadedConnect over MCP and the loop closes: your AI reads what you know mid-conversation and writes what you learn back, so every session starts smarter than the last.
Your link is private and unguessable. During the beta, hosted demo brains auto-delete after 14 days of inactivity; a brain on your own machine is yours for good. For a fully private or on-prem deployment, talk to us.