Work you repeat
Teach it by showing an input and the output you expect. After one explicit example, it can repeat that transformation whenever the same input appears again.
System Two thinking for your agent.
Recently, TypeSafe AI introduced Jev and its System One role in multi-model agent design, opening up a different way to think about how agents work. Today, we want to introduce you to Conceiver, a System Two counterpart to Jev. Conceiver learns from experience and brings what it knows to the task at hand. It can predict intermediate steps or supply relevant information, backed by evidence you can inspect. When it cannot support an answer, it abstains.
The agent wrapper
It learns your workflows so you can save tokens on repeated model calls.
The wrapper, included in the Developer plan and above, learns from your agent’s requests and tool calls. After observing a request handled the same way twice, it can recognize a matching request and execute the learned tool call directly, bypassing the LLM API call and saving the tokens you would have spent.
›
This demonstration runs under controlled test conditions. Results may vary with your workflow, tools, examples, and real-world operating conditions. The agent wrapper is included with the Developer plan.
System One and System Two
A deliberate partner for your LLM.
Reflexive models like Jev give you a fast response. Conceiver handles the other half. It works with your LLM, recalls evidence from what you've done before, weighs that evidence, and records where every decision came from. It learns from your examples and from what happens when it acts, so it gets better at your work over time. It is not an LLM. It has no weights and doesn't use backpropagation, and it isn't meant to replace your LLM.
Messy data
You can teach it data transformations by showing it unorganized input and the organized output you expect. For example, it can learn to sort and relabel a large, messy file system into organized data sources.
Your agent harness
If you run agents in a harness, the agent wrapper lets them learn from what they've already done. When it recognizes a supported step, it may be able to handle it without calling the agent's LLM.
How it decides
You can see why it answered.
It answers, or it tells you it can't.
If it has enough evidence, you get an answer. If it doesn't, you get UNRESOLVED together with the relevant evidence it has. Your application can use that evidence for the next decision, such as passing it to an agent's LLM or a person. Conceiver won't make up an answer to fill the gap.
file invoice from acme
ANSWER
/finance/vendors/acme/Comes with the evidence it usedfile invoice from initech
UNRESOLVED
There isn't enough evidence to select a destination.Relevant evidence is included for downstream decision-making
There isn't enough evidence to select a destination.Relevant evidence is included for downstream decision-making
You can read its reasoning.
Every decision comes with a short Markdown note you can read. It shows you where the answer came from, which examples backed it up, and which examples pointed the other way. That full record is what we mean when we say the provenance is lossless.
Examples become evidence.
A first observation supports a hypothesis. A second independent example establishes confidence in the reusable pattern. A third raises it to high confidence. Conceiver reports these distinctions through evidence levels, not probability percentages.
Hypothesis
One supporting example.
Confident
Two independent supporting examples.
High
Three or more independent supporting examples.
Contradictions lower its confidence. A multi-step answer is only as strong as its least-supported step. When answers disagree, Conceiver abstains instead of choosing one and provides the available evidence for downstream decision-making. The API returns ANSWER or UNRESOLVED with the evidence level, so you choose how much evidence you need before you act.
Your teaching outlasts the instance.
Everything you teach is saved to your training sequence. You can teach that sequence to a new Conceiver in seconds, giving it the same capabilities it had when you left off. Releasing an instance clears its working context, not the saved sequence.
An earlier research demonstration
Chess with no GPU and no LLM.
We ran 50 Conceivers at once against Stockfish at level 0. As they played, they all contributed what they learned to one shared knowledge store. The demonstration used hand-tailored chess rules that the base Conceiver does not include. You can download the separately loaded research starter model built from the initial games. It does not contain the later 1,000-game match records.
Conceivers at once
50
Each one added what it learned to the shared knowledge store.
Games against Stockfish
1,000
Every game was against Stockfish at level 0.
Games we taught first
400games
We started them with about 400 games, and we taught them the legal moves for every piece.
Draws
2
We used no GPU, no LLM inference and no LLM API calls.
What you get
What you get, and what it costs.
Conceiver starts with a Unicode lexicon. You teach it the vocabulary and transformations it needs for your workflows. All prices are in US dollars.
What you get on day one
Conceiver already knows Unicode characters and a small vocabulary. We kept its vocabulary sparse, so its own words won't clash with your workflow or count as evidence against it. You can teach it how to change natural language and how to handle structures nested inside other structures. Support for raw JSON syntax is in active development.
ConceiverYou use it through the Toolving API.
SDKYou teach, ask, save and restore from your own code.
Workflow examplesThree workflows you can copy, with the code to run them.
Workflow WorkshopYou're entered in the monthly workshop when you sign up.
// API usage, one shared credit pool credit price $20 per million credits prediction or retrieval 1 credit per successful operation episodic-memory query 2 credits per successful operation teaching 5 credits per successful operation inputs and evidence no additional charge minimum purchase $5 for 250,000 credits free signup grant 250,000 credits, once
Usage comes from one shared credit pool. A standard prediction or retrieval costs 1 credit. A memory-enabled query that uses contextual evidence and can recall full episodes costs 2 credits. Teaching costs 5 credits. You can prepay for credits or choose a workspace plan for storage and additional features.
We recommend adding storage before you use episodic memory. Without storage, your Conceiver instance and its episodic memories are ephemeral and disappear when that instance retires. With storage, its memories and learned capabilities are saved as an artifact that you can load into a fresh session.
Workspace plans
Free
$0 a month. You don't get storage on this plan, so upgrade to Developer or Pro to keep what you teach.
Developer Agent wrapper
$29 a month or $290 a year. You get 2 GB of storage and the agent wrapper.
Pro
$49 a month or $490 a year. You get 10 GB of storage.
If you need more storage, it costs $0.50 per GB-month, prorated, and you approve it through Stripe first. On the Free plan, you need to upgrade before you can add storage.
You get your 250,000 signup credits once. They don't refill, there's no monthly free top-up, and they don't roll over. Now and then we may hand out promotional credits.
Try Conceiver on your workflow.
When you sign up for a free account, you're entered in our monthly Workflow Workshop. Each month we pick one person's workflow problem to automate. Over seven days we write down how you work, find the spot where automation helps you most, show you how Conceiver works, and build the solution with you. By the end, you'll know how to take on bigger real-world problems yourself.
We record everything. Hydra-Dynamix reserves the right to use the footage in promotional material, including a tutorial video that will come with open-source code for the solution.