How Multilo remembers your project

Every new chat in Multilo starts with a fresh context window — it doesn’t carry the conversation from earlier chats. Two things carry what matters from one chat to the next:

• Project instructions: files you write once — AGENTS.md, voice.md, and integrity.md — that tell the AI how to work on this project.
• Agent memory: short notes an agent saves when you ask it to remember something.

This page covers how to:

• write and organize project instructions in AGENTS.md;
• set your voice and integrity rules;
• save facts to agent memory and choose their scope;
• view and edit what the AI knows;
• troubleshoot when instructions aren’t followed.

Instructions vs memory

The two work together. Instructions are rules you decide on; memory is facts that come up as you work.

Project instructions (AGENTS.md)

AGENTS.md is a plain Markdown file in your project’s hidden .multilo/ folder. It holds the conventions, terminology, and rules for this project, and Multilo loads it into chat and every agent so they follow your project’s guidelines without being told each time.

When to add to AGENTS.md

Treat AGENTS.md as the place you write down what you’d otherwise explain again. Add to it when:

• the AI makes the same mistake a second time;
• your supervisor or a reviewer points out something the AI should have known;
• you type the same correction you typed in the last chat;
• a new co-author would need the same background to help.

Keep it to things that matter in every chat. A step-by-step routine you run now and then is better as a custom agent or a skill, which load only when you use them.

Where instruction files live

All of these belong to one project. Open another project and it has its own .multilo/ folder, with its own rules.

Set up AGENTS.md

Open your project folder, show hidden files (View → Show → Hidden items in Windows File Explorer, or Cmd+Shift+. in the macOS Finder), and open .multilo/AGENTS.md in any text editor. If it isn’t there yet, create it. There is no required format — headings and short bullet points work best:

# Style
- British spelling; APA 7 for citations, with page numbers for direct quotes.
- Call participants "respondents", never "subjects".
- Keep paragraphs under 150 words.

# Project
- Research question: how does remote work affect trust in small teams?
- Chapter 4 is final. Don't edit it unless I ask.

# Sources
- Prefer peer-reviewed sources from my library over web sources.

Write effective instructions

Write instructions that are concrete enough to check:

• “Use British spelling” instead of “write properly”
• “Cite in APA 7, with page numbers for direct quotes” instead of “cite correctly”
• “Chapter 4 is final; don’t edit it unless I ask” instead of “be careful with finished parts”

And keep the file easy to follow:

• Short: for each line, ask “would removing this cause a mistake?” If not, cut it. Important rules get lost in a long file.
• Organized: group related rules under headings, with one rule per bullet.
• Consistent: if two instructions disagree — in AGENTS.md, voice.md, integrity.md, or a memory — the AI may follow either. Review them now and then and remove anything out of date.

For more on what to include and what to leave out, see Best practices.

Share instructions with co-authors

Because .multilo/ is plain files inside the project folder, your instructions travel with the project. Commit it to Git and co-authors who open the project get the same rules. Multilo’s own working data in that folder, such as the search index, rebuilds itself and is kept out of Git, so you don’t need to manage it.

Voice and integrity

Two more instruction files shape everything the AI writes in a project:

• voice.md describes how you write — your sentence rhythm, register, and habits — so drafts and edits sound like you. Multilo can learn it from samples of your own work.
• integrity.md holds your research-integrity rules: support claims with real sources, never fabricate a citation, and be honest about what can’t be verified. With strict mode on, the AI refuses to invent a citation and tells you when it can’t find a real source.

# Research integrity
- Cite only real sources from the project library.
- Never invent authors, journals, years, or page numbers.
- Say clearly when a claim can't be verified.

Set both from Settings, or when you create a project. They apply to chat, inline suggestions, and every agent, including the Full Draft Writer. See author voice & integrity.

Agent memory

Agent memory lets you tell Multilo something once and have it there next time — a preference, a decision, a detail about your project — so you don’t repeat yourself in every chat.

Ask an agent to remember

Just say what to remember:

Remember that my supervisor prefers short paragraphs and active voice.

Remember that we decided to leave the 2019 pilot data out of the analysis.

Each one is saved as a small, readable note. To see what’s saved, ask:

What do you remember about this project?

Project and global memory

How memory works

• You decide what’s saved. Agents only remember what you tell them to.
• Agents recall it on their own. When an agent works, it brings in the memories relevant to the task.
• It stays on your computer. Every memory is a local file you can read, edit, or delete.

View and edit

• Agent memory panel: review, edit, and delete memories from Settings → Agent memory.
• The files themselves: open the project’s .multilo/ folder and edit AGENTS.md, voice.md, integrity.md, or any note in memory/ like any text file.
• In chat: ask “What do you remember about this project?” to see what agents can recall.

See agent memory and the .multilo/ folder for the details of each file.

Troubleshoot

The AI isn’t following my AGENTS.md

• Check that the file is in the .multilo/ folder of the project you have open. Agents use the active project’s files.
• Make the instruction more specific. “Use British spelling” works better than “write properly”.
• Look for a conflicting rule in voice.md, integrity.md, or a saved memory, and remove one of them.
• Shorten the file. In a long file, the rules that matter get lost.
• After editing, start a new chat so the updated file is used from the start.

I don’t know what’s been remembered

Open Settings → Agent memory, or ask in chat: “What do you remember about this project?” Everything is a plain file you can read.

A memory is wrong or out of date

Edit or delete it in Settings → Agent memory, or delete its note from .multilo/memory/. Old memories that disagree with your current rules are a common cause of confusing answers.

An instruction got lost in a long chat

In a long chat, Multilo compacts older messages to make room, and detail from early on can be condensed. If the instruction was only typed in chat, put it in AGENTS.md — the files apply to every chat, however long.

A preference shows up in another project

Projects are kept separate: one project’s files, library, and project memory never reach another. Only global memory follows you everywhere. If a preference belongs to one project only, delete it from global memory and save it as a project memory or in that project’s AGENTS.md.

Related