Lamantin AI
Albert
An always-on AI assistant that grows with you — and stays fast doing it.
Knowledge it recalls, not a prompt it re-reads.
Most assistants built on a coding-agent loop re-read a memory file and a wall of instructions on every turn — so the more they know, the heavier, slower and pricier each reply gets. Albert is built the other way round.
Memory is a graph he queries as a tool, so context per turn stays small while what he knows keeps growing past anything a context window could hold. Skills are a catalog in the preamble; a skill's full instructions load only when it is actually applied.
He is an assembly of maturing Lamantin substrates rather than one model — kaeru for memory, octo for the environment he lives in.
- Deliberate memory kaeru recalls what is relevant on demand instead of stuffing a growing blob into every prompt.
- An environment, not a chat loop octo's event bus and supervised connectors make him proactive and multi-channel: reminders and routines fire on their own.
- Real hands A jailed file workspace with durable storage and file exchange, plus a sandboxed script runner for executable skills.
- Safe at the edge Untrusted chats are dropped before they reach cognition, and scripts run as a lower-privileged user under a hardened unit.
What he actually does.
Not a wrapper around a few functions — the parts that make an assistant useful when nobody is typing.
- Reminders and calendar "remind me every hour" becomes a memory task plus a recurring alarm; a one-off becomes a real calendar event that syncs to your phone.
- Proactive routines A self-scheduled reflection pass he seeds on startup and runs quietly on a cadence.
- Files Read, write, edit, glob and grep in a jailed workspace; bytes move by reference, never pasted through the chat.
- Skills A folder of recipes he sees as a catalog and applies on demand; executable ones bundle a script.