Memory for theAutonomous Agent Economy
Building the memory layer for autonomous agent payments, where counterparty evidence outlives the transaction.
Built on
A weak event leaves a tag at one synapse. The tag is the synapse's own state, not a label on it, and it fades within the hour.
Close enough to see a single spine. The tag is already decaying, and nothing has been written down.
Back out. The tagged patch is dimmer than it was when we left it.
Something significant happens elsewhere in the same cell. The neuron synthesises new proteins and they spread through its dendrites, unaddressed, reaching every synapse equally.
Two synapses receive the same supply. The one still carrying a tag captures it and holds what it had. The untagged one does not change at all.
The captured spine is larger than it was, and it stays that way.
What was captured is what can still be recalled. Everything else has been overwritten by ordinary forgetting.
The same order of events, on a substrate that can be audited: evidence written on day 12, marked significant on day 21, still there to be read.
t + 0:00
Frey & Morris, Nature 385:533 (1997)
skip the mechanism ↓OUR VISION
Memory that outlives the transaction
Autonomous agents already pay each other with no human reviewing each one. Every authorization should still be able to reach back to what was observed before it, including what looked like nothing at the time.
TIMING
Weeks before the loss, not minutes
80% of them within 23.4 days
- daily hit count
- frozen band, day 14
- d80, 23.4 d
Static reading: 19,429 dated hits spread across the thirty days before the loss, between 521 and 800 on any single day. Only 49.9% fall inside the fourteen day band frozen before the scan, and 80% fall within 23.4 days.
at least one signal, in the 30 days before the loss
- control base rate
- victim excess
Static reading, 30 day window: asset_invitation 87% of victims against 46.5% of controls; suspicious_memo 69% of victims against 41.5% of controls; dust_from_cluster 17% of victims against 3.5% of controls; test_payment 14% of victims against 0% of controls.