Institutional memory rides in the cab
What your operation knows lives in radio calls, group texts, shift-change conversations, and the heads of your most senior people. It is written down constantly — and retrievable never.
So the same questions go unanswered every week: Has this crossing gate failed before? What usually goes wrong on the Springfield segment? What did the last inspector flag on this car? The answer exists. It's just three months deep in a message thread nobody can search by place.
And when a dispatcher retires or a driver moves on, twenty years of "which dock needs the short trailer" walks out the door with them.
Memory with coordinates and a clock
Aivery is memory infrastructure for operations. Field notes and chat messages become structured memories automatically — each one stamped with when it happened and where (from the device that sent it, no forms, no dropdowns).
Then anyone — or any AI agent you run — can ask questions the way an ops person actually thinks: by facility, by corridor, by time window.
Responses above are real output from Aivery's spatial test runs — not mockups.
No map service knows your Dock 7. Your operation does.
Google can geocode "Springfield, IL." It cannot geocode the Springfield switching station, the north staging yard, or Dock 7 at the Peoria terminal — the places your people actually talk about.
Aivery learns them. When crews mention a facility while working at it, the name binds to its coordinates automatically. Your operational geography assembles itself — no configuration, no GIS project, no data entry.
| Facility name | Coordinates | Mentions | Source |
|---|---|---|---|
| Vulcan Staging Yard | 41.6100, −87.3350 | 3 | learned |
| Proviso Yard | 41.8897, −87.8750 | 2 | learned |
| Springfield switching station | 39.7837, −89.6437 | 2 | learned |
Public places resolve against a 250,496-entry gazetteer bundled with the system — including 15,000+ US rail yards, stations, ports, and airports. Fully offline: no memory content ever touches an external geocoding service.
From message to memory to answer
Capture
Crews send notes the way they already do — chat, app, or API. Device GPS and time ride along automatically.
Extract
An LLM distills each message into discrete, durable facts. Verbatim details — names, weights, car numbers — are preserved.
Enrich
People, equipment, and places are indexed; facility names bind to coordinates as the corpus grows.
Recall
Query by radius, corridor, facility name, or time window — combined with meaning-based search across everything.
Runs in our cloud or entirely on your infrastructure. PostgreSQL + PostGIS at the core — boring, auditable, yours.
Tested like infrastructure, not a demo
* LOCOMO long-conversation memory benchmark, matched answer format: LLM-judge 0.758 vs 0.638, F1 0.461 vs 0.406, BLEU 0.384 vs 0.360. Full methodology in our technical paper.
Where place-aware memory pays
Recurring faults by location
"Third gate failure at this crossing in two months" stops being tribal knowledge and becomes a query.
Crew & shift handoffs
The next crew inherits everything logged on this route, this yard, this car — not whatever fit in a five-minute conversation.
Corridor incident review
Reconstruct any run end-to-end, in order, from the memories written along it. Audit trails assemble themselves.
Customer-site knowledge
Every delivery quirk — the dock that needs the short trailer, the gate code, the receiver who signs late — retained per site, forever.
Maintenance history in place
Ask what's been flagged within 500 meters of where your inspector is standing, before they climb down.
Agent-ready operations
AI dispatch and planning agents get the same recall your people do — through one API.
Founding Partner pilots — now open
We're onboarding a small number of logistics operators as Founding Partners: a scoped pilot on one route, yard, or fleet segment, running against your real operational traffic.
- Working deployment on your data in weeks, cloud or self-hosted
- Direct access to the founding team; your workflows shape the roadmap
- Founding Partner pricing, locked in