AI agent project memory

AI agent project memory is the durable record of what a project is, what has been decided and what has already been tried, kept somewhere other than the chat window. Most developers meet the problem the same way. You spend an hour with an agent establishing conventions, rejecting one approach and choosing a file layout, then open a fresh session the next morning and explain all of it again. Chat history is a transcript, not a memory. It is ordered by time, hard to search by purpose, tied to one tool, and quietly truncated as the context window fills. For a solo developer this costs minutes per session. For a producer coordinating several agent sessions across several projects it costs the plot. The useful question is not how to make the model remember more. It is where the facts that should survive a session live, who is allowed to edit them, and how an agent receives exactly the relevant slice when it starts work.

Chat history is a log, not a memory

A conversation with an agent records everything in the order it happened: the false starts, the corrections, the pasted error output. That is useful as an audit trail and poor as a source of truth. The decision that matters, for example that the importer must never touch the shared config, sits at message 41 of a thread nobody will reread. When the thread grows past the model's window, the early messages are the first to be summarized or dropped, and those are usually the ones that set the ground rules. Threads are also private to the tool that produced them. A conversation in one CLI is invisible to a second CLI, to a teammate, and to you three weeks later when you try to remember why a module looks the way it does. Treating the transcript as memory means every new session starts by paying a re-explanation tax, and every re-explanation drifts a little from the last one.

What durable project memory actually contains

Durable memory is a small set of records with stable identifiers. A work item describes one outcome, its priority, its dependencies, the files it is allowed to touch and how you will know it is finished. A decision record states what was chosen and why, so the next session does not relitigate it. A history of attempts says what ran, what the result was and what changed. The key property is that each record can be found by something other than a timestamp: an identifier, a topic, a file, a status. Good memory is also bounded. Feeding an agent the entire project history is not memory, it is noise. The goal is to assemble the few items the current task needs, say what was left out, and let the agent work from facts instead of from whatever it inferred from the last message you typed.

Where the record lives in Liquid Brain

Liquid Brain, a Windows desktop application called the Desk, treats the project card as the unit of memory. You create cards, edit their details and track them through the board [C1a], and you can find them again by ID, title, domain, scope or artifact [N01]. Cards carry assignee, reviewer, due date, blocker and dependency details [N03]. When an agent is about to work, task context can be assembled with dependencies and relevant notes, and anything omitted is disclosed rather than hidden [N11]. The outcome of agent work is kept with the card: recorded attempts, results and reported usage [C4]. Recorded evidence can also be checked for freshness, so you can see whether it still matches its supporting files [C6]. A matching hash tells you the bytes are unchanged, not that the check was adequate, so the freshness signal supports your judgment rather than replacing it.

QUESTIONS

Is a long context window the same as project memory?

No. A large window holds more of the current conversation, but it resets when the session ends and it has no structure. Project memory persists across sessions, has identifiers you can search, and lets you choose what an agent sees at the start of a task.

How do I stop an AI agent forgetting decisions between sessions?

Write the decision down in a record the next session will actually read: a work item, a decision note or a project instruction file. Then make loading it part of starting work. In Liquid Brain, task context pulls in dependencies and relevant notes, with omissions disclosed [N11].

Does Liquid Brain store my project memory in the cloud?

The facts on file describe a Windows desktop Desk that runs on your computer [C10]. Nothing in the approved material describes a hosted storage service, so do not assume one. Release acceptance is still in progress and the product is not purchasable yet.

RELATED GUIDES

• What is a project management harness?
• How to create a project card in Liquid Brain Desk
• How to find and organize work across projects

RELATED HARNESS PAGES

• project cards vs chat threads Project cards vs chat threads: when a conversation with an AI agent is enough, and when a structured card record is the better home for the work.
• CLAUDE.md alternatives CLAUDE.md alternatives compared: what a project instruction file does well, where it stops, and what a project harness like Liquid Brain adds.
• local AI agent workflow A local AI agent workflow keeps project records and agent runs on your own computer. Here is what that gives you and what it costs.