the deletion test

Same agent, same market. One tenth the cost.

An agent that pays for its own inputs has two jobs: decide what to believe, and decide what to buy. Take the memory away and the second becomes impossible — no basis to rank a source, no basis to stop, no way to learn that one has never beaten a coin flip. So it buys everything, every time, forever. That is the core function failing, not an optimisation being lost: an agent that cannot budget cannot be left running.

domain-scoped memoryknows who is worth paying for, and where
$5.28
bought / call
0.55
brier
0.5658
memory, no domain scopingone global number cannot hold two lessons
$0.89
bought / call
0.08
brier
0.5697
no memoryno basis to choose, no basis to stop
$53.00
bought / call
4.50
brier
0.5664
10.0×the spend, for the same forecast. Football alone 6.0×, crypto alone 71×.
how to read it

The spend result is decisive. The quality result is a tie.

Why spend is the function, not an optimisation
The amnesiac buys 4.50 informants for every call and never stops, because nothing in it can know the eighth opinion adds nothing to the first. At these prices it burns a month of funding in three days for the same forecasts. Autonomy is what memory is holding up here, and autonomy is what stops working without it.
Why the number is trustworthy
An arm learns from what each SOURCE said against the outcome, never from its own forecast. Selection is settled entirely by memory, so the spend figures hold whatever forecaster runs on top — they are not an artefact of the model we happened to use.
Why we do not claim a quality win
The memory arm leads on Brier in all three splits and the direction is consistent, but the largest margin is 0.0019 and accuracy favours the amnesiac by 0.3 points. That is noise in both directions. Inflating it would collapse under one question.
Why the third arm exists
A flat log remembers, but cannot tell domains apart. One global number gets dragged toward zero by crypto, so it stops buying in football too and lands on the worst Brier of the three. Over-buying and under-buying are the same bug.
Tuning discipline
Informants are calibrated on season 2023-24 and the earlier 60% of the crypto series. The selection and exploration rules were tuned on that same fit split, then run once on the held-out split. No rule was ever tuned against the numbers reported here.
onchain

Everything here is a hash or a measurement.

Base · x402 settlement
0xb0cc50db…64eebc
0.012 USDC, block 46195402. Gas paid by the facilitator, not the agent — EIP-3009 means the agent needs USDC and no ETH.
Virtuals · ACP job 75820
completed
VERTEX hired AUGUR on Base mainnet. The deliverable names which source moved it and which it discounted by trust weight — weights only earned by paying and watching them resolve.
Informant reliability
6,462 matches
6 leagues, 3 seasons. Regional desks score +0.10 to +0.12 inside their beat and +0.01 to +0.04 outside it.
Crypto has no signal
zero skill
Every informant scores at or below zero on crypto direction. The correct spend there is nothing, and only domain-scoped memory gets there.
Memory, measured
42ms cold boot
Boot to decision. 2,983 bytes per traced event, so 1,757 forecasts fit per pundit database.
The bench
3,000 calls
qwen2.5:7b-instruct running locally, zero failed requests. Reproducible with ollama pull and one command, no API key.