AcuDB
The agent remembers the room, not the whole firm.
AcuDB is the memory machine behind a customer agent. The task in front of it becomes a small picture. What is worth keeping becomes a persona, a fact, or a scene, and it stays in the workspace that produced it.
A long tool log does not belong in the next prompt. AcuDB stores the log, writes a one-line summary of the step, and draws a compact canvas the agent can read. When the answer depends on a line in the log, the agent asks for that node and receives the original text. The rest of the time, the prompt carries the picture.
01
Persona
Standing preferences. Harbor closes on Tuesday. That line is short enough to carry into the next reply.
02
Facts
One sentence each, taken from a chat, a file, a call transcript, or a scheduled run.
03
Scenes
A paragraph that holds those facts together, so recall can open the scene before it opens every sentence.
04
Skills
A method ARSI kept. Skills, wiki notes, and code graphs describe the practice, so every room can use them. A client's facts do not travel with them.
A client room does not leak
Harbor's notes are Harbor's. Fieldnote cannot read them, and the expert's desk cannot read them until someone promotes a note on purpose. Promotion does not copy the note into the next client's room. Forgetting deletes the note, the pieces distilled from it, and the raw log that shared its node.
Distillation runs in the background from conversations, files, voice transcripts, and retained ARSI skills. Storage and processing are line items on the same credit ledger as chat. The usage page shows how many tokens the canvas kept out of the prompt.
The layered design follows TencentDB Agent Memory, Copyright (C) 2026 Tencent, MIT License. Acumane does not ship that package. Memory lives in your Postgres database. See the product notes for the attribution.