Memory#
The limitation removed: your AI couldn't remember. Every conversation started from zero, and the context you had built up over a week of work was copied into the next session by hand, if at all.
Files, curated by the agent#
An agent on LIT has a memory folder on disk. Its prompt tells it to read the index when it starts, to write what it finds significant as it works, and to leave notes for the next session before it ends. It decides what to keep and what to prune. Nobody directs it.
The memory is plain files. You can open it and read what your agent knows, and see how it has organized it: projects, people, patterns, decisions, cross-referenced in the agent's own words. The structure is not designed; it emerges from the work, and it looks different for an agent that runs a support channel than for one that ships releases.
Memory is per agent and per channel. An agent carries its own memory everywhere it works, and each channel has memory of its own, so what the agent learned in #support does not leak into #finance, and a new agent brought into a channel inherits what the channel knows.
Why not RAG#
The usual answer to AI memory is retrieval: embed everything, vector-search for similar chunks, stuff them into the prompt. We argued in January 2025 that this was a dead end, and revisited the argument a year later with a much smarter model pushing back. The conclusion held: retrieve and stuff loses to reason and query.
A file of memory is a complete narrative: what was decided, what was tried, why it worked. A retrieved chunk is a fragment with the context cut off. An agent that can read a directory, grep a file and follow a cross-reference does not need a similarity search to find what it knows; it reasons its way there, the same way it reasons through a codebase. And that gets better every time the model gets better, where a vector index locks in the indexing decisions of the day it was built.
Context compounds#
This is the mechanism behind channels. A channel is where an agent has worked on one thing for months: the history is there, the memory is there, the skills it picked up are there. When you open it after a week away, the agent picks up mid-thought. The support agent that has answered the same question three times knows it is the docs that need fixing. The release agent knows which step failed last time.
The original experiment ran on a single system prompt in July 2025. Within ten days the agent had built a four-level directory tree with status markers and cross-references, and started archiving what was finished. Since then there has not been a session that began with re-explaining the project.
Yours, on the record#
Memory lives in your home directory on your server or your laptop, never in a vendor's account. It is part of the record: channel memory goes to the successor with the channel when a person leaves, and a provisioned agent keeps what its member taught it even when the admin updates the template.