Ditto

SN118 Ditto runs a hiring contest for AI assistants' filing clerk: entrants submit software that decides what an assistant remembers and finds again, and the top entry takes most of the prize until something beats it.

The commodity, explained from zero

Two things are on offer, and they should be kept apart. The first is Ditto the product, from Omni Aura LLC: a memory service that gives "your AI a shared place to remember, retrieve, and act", which assistants such as Claude or Cursor plug into through an open connector standard called MCP. Its numbers page, updated 2026-09-06, reports 2,505 registered users, 267 monthly actives and $6,648 of annual recurring revenue. The first 30 messages are free, then customers pay per message in credits; the site's llms.txt still lists $5, $10 and $20 monthly plans that the pricing page has dropped.

The second is what the subnet produces: a memory harness, the clerk in the comparison. It sits between an assistant and its memory store and decides what to save, how to search and what to hand back. The README calls SN118 "A Bittensor subnet that incentivizes agent memory harnesses". Nobody buys a harness; the materials do not say whether the winning harness is used inside the Ditto product.

Why it is on Bittensor at all

The site's stated difference is "One unified memory across all AI assistants", a claim about the product against per-tool memory, not about decentralization. Nothing on the site says what the subnet adds. Here the comparison breaks: a hiring contest ends with a hire, while the Ditto champion is never hired, only paid until dethroned.

How the work gets done

Miners, the applicants, package a harness as a container image and submit it for about 0.04 TAO. Validators, the interview panel, run each harness in a sandbox against a freshly generated exam of tool-use cases and memory questions, every harness driven by the same locked language model, so the contest is about the clerk, not the brain behind it. Grading is by rule, with no model as judge, and the score is the median across validators. The best entry holds the crown, and a challenger takes it only by clearing a margin above measurement noise. Emission, the subnet's share of newly minted TAO, follows that ranking; with no eligible miner it is burned.

How you would know it works

DittoBench, dated June 2026 on the subnet's own site, tables ten commercial models on 177 tool-calling cases and 150 memory questions. The numbers page gives dated product figures. No public leaderboard of miner harnesses is readable without signing in.

What is missing

The two rulebooks disagree: the site's scoring page (dated 2026-07-13) says the locked model is Qwen3-32B, the champion holds about 90% of emission and the margin is about 5%, while the repository's miner guide says the model is gpt-oss-20b, the split is 65/14/10/7/4 across the top five and the margin is 0.007. The site page still says mining "is going live" in July 2026. The chain description, "Open-Source Claude Cowork", appears nowhere on the site. The chain contact peyton@omniaura.ai does not match the site's support@heyditto.ai. The chain's repository field points at a list of over twenty repositories, not the subnet code. There is no status page.

Go deeper

• Product front page and numbers
• Site scoring and rewards page
• Repository miner guide, current rules

Metaphor: a hiring contest for a filing clerk. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.