Agents that learn from every conversation, answer in realtime.
One multi-model memory layer behind whichever model you run, trainable on your own data. Preferences learned from real conversations, returned in real time, and benchmarked on the state-of-the-art memory suites.
{
"user": "u_8213",
"query": "Can you draft an email to john",
"context": [
"prefers short replies, no preamble",
"john works in insurance claims",
"has a casual tone with john",
"always asks for the source"
]
}"Hey John — file attached. Source: the Aug 14 export."One call, before your prompt. No fine-tuning, no vector store to run.
{
"user": "u_4417",
"query": "Add retries to the upload handler",
"context": [
"explicit error types, never a bare except",
"no new dependencies without asking",
"ships a test with every fix",
"prefers small diffs"
]
}"Bounded retry with an explicit UploadTimeout, plus a test. No new deps."Same call, another user. The context is theirs, not a prompt you maintain.
What changes
Three numbers, on public benchmarks.
Memory is easy to claim and hard to measure. These are ours, on a public benchmark, with the method written down.
Our own evaluation runs, scored with an LLM-as-judge rubric. Method and prompts available on request.
How it works
Two calls, and it starts remembering.
No fine-tuning, no vector database to operate, no migration. It sits beside the model you already use.
The benchmark
What it costs as the conversation grows
Two charts on one timeline: recall accuracy above, cost below, indexed to each system's own week one. Ours dips while it learns you, then recovers; context-stuffing decays and never comes back. Illustrative until the task behind a dollar figure is defined.
Recall accuracy
How much of what a user told you is still recalled correctly as the conversation grows. Context-stuffing decays because the window fills and the earliest turns fall out of it; a memory layer keeps what mattered and recovers.
Cost, multiple of week one
What the same conversation costs to serve, indexed to each system’s own week one and drawn on a log scale. Re-sending the whole history on every turn compounds; storing it once and retrieving what is relevant does not.
What you can build
Things people are using us for.
Anything that talks to the same person twice. One REST API, whichever model you're on — OpenAI, Anthropic, Google, or your own.
{{ u.title }}
{{ u.description }}
Example: {{ u.scenario }}
What it learns:
Outcome:
{{ u.outcome }}
{{ u.apiFlow }}
How it works: Send user interactions to our API, retrieve personalized context for each session, and inject it into your AI prompts. Your application learns and improves continuously.
We’ve built state-of-the-art ML models before.
SapientPriors is a research group working on stateful memory, context management and continual learning. The API is the part of that work that's ready for other people to use.
Where the team is from
Get Beta Access
Your first memory call is 30 minutes away.
Bring one assistant that keeps forgetting. We'll wire the memory layer into it on the call and you'll leave with a before-and-after you can show your team.
What beta partners get
{{ i.title }}
{{ i.detail }}
Prefer email, or LinkedIn — contact the founders directly
Request received.
{{ sentBody }}