It answers in about a second and a half.

A small model reading a compiled, on-device memory — where feeding a much larger model your whole history takes tens of seconds, and sends it off your machine.

The Kaineros desktop app answering a PersonaMem-v2 persona: a personalised, second-person recommendation that draws on your late-night binge-watching habit and pollen allergy; the three pages it read listed at right; a 1.9-second response time; and a low-poly face beside the chat. — Answering a PersonaMem-v2 persona in **1.9 seconds** — reading three compiled pages (listed at right), phrased directly to you, with its reasons shown beneath the answer.

We measured Kaineros against PersonaMem, a benchmark for tracking a user across a long, evolving conversation, three ways — the same questions, scored the same way:

Kaineros — a small model reading the compiled wiki, on your machine. The slow compile happens once, off the clock.
The same small model, raw history — handed the full raw conversation instead of the wiki. Isolates what the wiki changes.
A much larger model, raw history — reading the whole conversation. A ceiling reference.

PersonaMem v2 — 5 personas, 129 questions
Setup	Reads	Accuracy	Response time
Kaineros	compiled wiki	0.41	~1.4 s
Small model	raw history	0.50	~28 s
Large model	raw history	0.67	~28 s

PersonaMem v1 — 5 personas, 49 questions
Setup	Reads	Accuracy	Response time
Kaineros	compiled wiki	0.61	~1.6 s
Small model	raw history	0.73	~23 s
Large model	raw history	0.78	~23 s

Response time is time-to-first-word, measured. The raw-history rows must read tens of thousands of tokens before the first word; Kaineros reads a handful of compiled pages.

Reading it honestly

It trades a little accuracy for speed, cost, and privacy.

Reading the raw history is more accurate at this length — and Kaineros doesn’t beat it there. What the compiled memory buys is the thing an assistant actually needs to be usable:

~15× faster to the first word — about 1.5 seconds, not 25. A model re-reading your whole history isn’t an assistant; it’s a spinner.
~20× fewer tokens per answer, on a small model — cheap enough to run locally.
Private. It reads a few pages from your own disk. Nothing pours your life into a context window.

And the gap is where the work now goes: teaching the fast search to reach the raw detail it needs, on demand, without giving up the speed.

Caveats

Small samples (129 and 49 questions, 5 personas each) — a slice, not the full benchmark.
Our own evaluation harness (PersonaMem’s data and format, our loader and scorer).
Kaineros is a two-part system — a large model compiles offline, a small one answers — measured against single models reading raw context.
One context length (32k). Every claim is checkable against the code.

← back to kaineros.com