The model forgets. The work doesn’t.
Synthesis is the continuity layer for AI applications. Every model you use works from the same record, and none of them writes to it directly. Each one proposes; the rules decide what commits, one change at a time; and nothing is approved or run if the facts changed after it was proposed.
Your AI isn’t forgetful. Your system is.
Every API call starts from nothing, and the moment you use more than one model, each one starts from a different nothing. Whatever the model knows about your customer, your rules and last week’s decisions, you rebuild for it on every call. Most apps do that in one of two ways, and both quietly lose work.
The whole thread, every note, every version of the policy. Costs climb, and when two facts disagree nothing tells the model which one is current.
Cheap and fast. Then Tuesday’s correction falls out of the window, and the model confidently repeats Monday’s number.
Scoped by permission, ranked by authority, checked for freshness, held to a budget, and written down so you can see exactly what the model saw.
One account. Three days. Three ways to brief the model.
A support agent gets a simple question on Thursday. Switch the context strategy and see what the model gets, and what it says back.
Every call runs on a contract.
Five steps run before the model sees anything and two run after it answers. Permission comes first, so nothing a caller shouldn’t see can reach the prompt by accident.
Scope
Who is calling, for which work. Permissions filter the records before anything is gathered.
Authority
Structured state wins in its own domain. When sources disagree, the conflict is surfaced instead of settled silently.
Freshness
Every fact carries its source, its time and what it replaced. Old facts stay findable but stop steering.
Assemble
Instructions, the runbook section in force, live state, relevant records and recent turns, inside a fixed budget.
Manifest
A record of exactly what went into the call, so a bad answer can be traced to a bad input.
Any model
Whichever model takes this call does the thinking. Synthesis holds the record every model shares.
Propose
Output is a proposal until it passes the rules: current state, authority, and approval where it’s required. Proposals from different models are settled one at a time, and nothing commits if the facts changed after it was proposed.
Receipt
Approved actions leave a receipt linking evidence, approval and effect, so they can be traced and recovered.
Every model. One record.
Synthesis started because we use multiple models on the same work every day, and they kept overwriting each other. So in Synthesis, no model writes the truth directly. Each one proposes; the contract decides what becomes current, one change at a time. Whichever model takes the next turn starts from the same facts.
Continuity across the business. Depth inside every stage.
Across, work moves from research to an offer, to a sale, to support, and the promises travel with it. Down each stage runs its own improvement loop: research a gap, propose a change, implement it within authority, support the result, then start again from what you learned.
| Loop ↓ · Stage → | Research | Build / offer | Sell | Support |
|---|---|---|---|---|
| Researchfind a gap | Scan a market for unmet need | Find where the offer falls short | Find who’s ready to buy | Spot the repeat ticket |
| Proposeshape a change | Frame a bet worth testing | Spec the change | Shape the offer for them | Draft a fix or runbook change |
| Implementact within authority | Run the smallest test | Build it | Quote within pricing rules | Ship it within authority |
| Supportobserve & maintain | Watch the signal over time | Keep it working | Track every promise made | Confirm the tickets stopped |
| → Work hands off left to right, with its records attached.↺ Each column loops back to its own Research. What Support learns starts the next round. | ||||
The handoff is where work gets lost. The record Sales made on Monday is the record Support reads on Thursday, with its source attached.
Each stage can investigate, propose and act inside its own authority, instead of only passing tickets down the line.
Today’s product is the layer underneath: one record, precise context and governed change.
Designed to sit in front of the tools you already trust.
Synthesis doesn’t replace your memory store or your workflow engine. It governs what passes between them and the model.
Good tools already exist for storing memories and for running durable workflows that pause for a human. Synthesis is designed to work with them. What it adds is the part teams usually hand-roll and rarely write down:
- Which source wins when the CRM, a runbook and yesterday’s chat disagree.
- What a caller may see, applied before the prompt exists.
- What a model may change, with output held as a proposal until it passes.
- A manifest and a receipt for every call, so mistakes can be traced and recovered.
Synthesis started as a practical problem. We use multiple models on the same work every day, and they kept losing or overwriting each other’s work. Synthesis is the layer that gives them one record: each model proposes, and the rules decide what becomes current.
Its first operating context is Marketomic, a web and automation studio that researches, builds, sells and supports its own work.
Tell us what keeps getting lost.
Early access opens to a small number of teams running more than one model across more than one kind of work. Your answer to the last question shapes who we talk to first.
Straight answers.
Can I use Synthesis today?
Not yet. It’s in development and running in its first operating context. There is no public SDK or hosted service. Join the waitlist to hear when early-access conversations open.
Is Synthesis an orchestrator?
No, and it works with the one you have. Your orchestrator decides what happens next. Synthesis decides what each model is shown and what it’s allowed to change, so every step starts from the same record and nothing commits on old facts.
Do the models have to come from one vendor?
No. Every model you use is briefed from the same records and writes back through the same rules. We use multiple models ourselves; that’s why it exists.
How is this different from LangGraph, Temporal or Inngest?
Those are strong at durable execution and pausing for human approval, and Synthesis is meant to work alongside them. Its job is what the model is given and what it’s allowed to write back: permission-filtered, authority-ranked context, proposals kept apart from facts, and receipts for every action.
How is this different from Mem0, Letta or Zep?
Those are memory systems, and some already handle timestamps, provenance and superseded facts well. A memory store can sit behind Synthesis. Synthesis focuses on the contract around memory: which source is authoritative right now, what the caller may see, and what an output has to pass before it becomes durable. Whether that beats building it yourself is something we’re measuring.
Does it stop hallucinations?
No. Models can still be wrong. Synthesis makes errors easier to catch and harder to persist: proposals stay separate from facts, sources stay one step away, and every call has a record of what the model actually saw.
Will it cut my token costs?
We don’t promise savings. A bounded context can help, but the result depends on your workload and should be measured against what you do today.
Who is it for first?
Small teams running real business work through more than one model or agent, such as research, delivery, sales and support, where a full enterprise platform is more than they need. The people running that work are who it ultimately serves.
Will it be open source?
Undecided. If the record, provenance and receipt formats prove reusable beyond their first application, publishing those formats is the likely first step. We won’t promise a license or a date before then.