The memory engine
One memory, with every agent reading from it. Each agent takes what it needs from the same record and puts back what it learned, so a handover between them costs nothing.
Each point is one recorded memory: four days of one operation.
Memory, not a knowledge base
Most AI tools in this industry search a knowledge base that you keep filling in. The memory engine keeps a record of what actually happened, and it grows as the work gets done.
It remembers what happened
Who called, when, what was done and what it cost, not only what someone wrote down in advance.
What it keeps
Facts that can be checked, what happened and who was involved, a summary of every person, property and contractor, and beliefs that change over time, each kept apart. Nobody writes memories by hand: the night’s notes, the contractor’s report and the owner’s email become memories on their own.
It knows what is still true
When something changes, the new fact wins. The old one is kept, with its date and the reason it changed.
What the research measured
The engine is built on published, open-source research from December 2025. On standard academic benchmarks, the paper reports answers that depend on a changed fact rising from 60.3% to 84.6% against its baseline, and answers that depend on when something happened rising from 31.6% to 79.7%. These are the paper’s results, not measurements of our own properties.
It can name its source
Every memory traces back to a plain text note, with when it happened and who was involved.
Why that matters
When an agent says the boiler was serviced on the fifteenth, it can show the note, the time and who was there. When it does not know, it has no partial match to dress up as an answer. Each note also carries a scope, so an internal discussion cannot surface in a reply to a guest.
The memory is the part you cannot buy back
Staff move on and take what they knew with them; the memory stays. It holds the record, not the judgement: when a conversation turns to money, blame or liability, a person still takes it.
Sources
- Latimer, C., Boschi, N., Neeser, A., Bartholomew, C., Srivastava, G., Wang, X. and Ramakrishnan, N. Building Agent Memory that Retains, Recalls, and Reflects. arXiv:2512.12818, 14 December 2025. arxiv.org/abs/2512.12818
- VentureBeat, report on the research and its open-source release. venturebeat.com