An open experiment in AI memory

Before it becomes
a memory.

Should your AI keep that thought?
We use Jev to test whether a memory is supported,
useful over time, and faithful to what was said.

Try it yourself ↗Explore the repo ↗

Open source · Recorded examples need no key

Measured, not promised.Loading recorded study…

Mean Jev response

—msClient-observed request time

Preview agreement

—%Against our expected labels

Intended memories saved

—Preview policy · unique cases

English evaluations

—One evaluation per distinct case

100 agent-authored synthetic cases; no independent human annotation. These are admission recommendations, not verified facts or stored memories.

The findings

A threshold changes what gets kept.

Read the experiment ↗

Same responses. Two policies.

Matches / 20

Baseline · 0.75Preview · 0.40

Loading repeatability results…

Every response time, kept.

Milliseconds

—Median

—95th percentile

—Slowest

First requestInterleaved categoriesLast request

Includes network time and the first request. No outliers removed. Not a throughput test or speed guarantee.

What still needs work

Loading observed limitations…

Earlier repeatability study ↗

The standard

Three questions.
One recommendation.

Supported?

Does the original passage actually support the proposed memory?

Useful later?

Will this matter beyond the current conversation?

Faithful?

Are proposals, conditions and uncertainty preserved?

Your turn

Try it yourself.

Start with a recorded case. Then test a memory of your own.

Save / Skip / Defer

Explore examplesReal recorded results. No key, no new API calls.

Evaluate your textRun locally with your TypeSafe key. One API call per evaluation.

START WITH A RECORDED CASE No key. No API call.

The evidence

The judgment

Waiting for a case

◇

A memory starts with evidence.

Choose a recorded example above,
or evaluate your own source and candidate.

Help us find the edge cases

What would your AI get wrong?

Try your own synthetic examples. Share the expected decision
and the actual result so others can reproduce it.

Run the lab locally ↗Share a case on GitHub ↗

Part of Cairn

Before storage comes a decision.

This lab explores what deserves to become a memory.
Cairn Memory is the memory project; this demo does not write to it.

Explore Cairn Memory ↗