Working with Memorable on Faster, Cheaper Agents
Voice is unforgiving about latency. There's no spinner to hide behind, no progress bar to watch. When an agent takes a beat too long, the person talking to it just hears silence, and silence is the one thing a voice product can't explain away.
Where our memory left off
That's most of what shapes our roadmap right now. Agents on OpenHome have gotten good at remembering people. Preferences, history, the things you've told them before, all of that carries across sessions and it works.
What we've been after is the next layer: agents that also remember how to do things, so a task they've completed a hundred times doesn't get rebuilt from scratch on the hundred and first. That rebuild is where the silence comes from. The more steps a request has, the longer the agent spends working out an approach it has already worked out before, and the longer the person waits to hear anything back. Reading a stored workflow takes milliseconds. Working one out means going back to the model at every step and waiting for each answer before the next one can start.
What Memorable does
That's the problem Memorable set out to solve, and it's why we brought them in. Most memory tools store what happened. Memorable stores how things get done.
It watches an agent's own successful runs and turns them into reusable workflows, so the path an agent worked out once becomes a path it can follow. Nothing gets fine-tuned, nothing is locked to a particular model, and a workflow describes what it's after rather than which tool it used, so it survives the kind of stack changes that happen constantly on a platform like ours.
Why prompt tokens stop scaling
The usual way to teach an agent a procedure is to write it into the prompt. It works, it's simple, and it's what nearly everyone does. It's also the thing that stops working first.
Every step of every procedure sits in the context window and gets paid for on every call, by every user, whether that particular run needs it or not. One agent doing this is a rounding error. A catalog of abilities running across tens of thousands of people is not.
When a procedure is already spelled out in the prompt, memory adds nothing on top. It's the same result at a higher cost. And the obvious alternative, trimming procedures out to save tokens, is worse than it sounds: with the procedure gone and no memory to fall back on, an agent doesn't degrade gracefully. It stops and asks what the routine even means.
So the choice was to pay for every procedure on every call forever, or find somewhere else to keep them.
Running it live
We ran it on our system earlier this month against eight real tasks, twice over. Once with procedural memory switched on, once without, same tasks both times. With it, all eight completed. Examples included both ordinary and multi-step actions: a reminder set by voice, which fired right on schedule. A flight from San Jose to Austin, booked start to finish off a stored workflow, with the window-seat preference applied along the way. And a movie night routine, lights and all, run straight from memory with nothing re-explained.
Testing showed this holding at larger scale: near-perfect completion with procedural memory, steady across repeat runs, and no wasted tool calls along the way.
Why this lowers cost, including for BYOK developers
Moving procedures out of the prompt is where that gets paid back, and it gets paid back twice. Leaner calls come off our own bill. But a lot of our developers bring their own keys and their own provider billing, and for them fewer tokens per call is money back in their pocket, not just ours.
What it opens up
Multi-step, multi-ability workflows have always been the hard thing to ship on a voice platform. They get easier when agents can build on what they've already done well, and the payoff shows up on both sides at once. Actions land faster, so there's less dead air between asking and hearing something back. And the procedures behind them cost less to run, which is what makes shipping more of them possible in the first place.
More from Memorable on how it works at memorable.sh. If you're building on OpenHome, come find us in Discord.
About Memorable
Memorable builds procedural memory for AI agents. Instead of storing what happened in a conversation, it stores how work gets done, turning an agent's successful runs into reusable workflows that survive model and tool changes. Learn more at memorable.sh.