Per-project instructions and memory

[TheNomadInOrbit]

MonoCode version

0.1.42

Operating system

macOS

OS version

macOS 26.6.2

Provider CLI

Codex

What happened

What I want

Each project should have its own set of instructions and its own memory, so a session started inside that project automatically carries the project's context.

Concretely: right-clicking a project should let me edit that project's instructions, the same way I can already set a custom logo for it today. Every new session opened in that project should pick those instructions up automatically.

Why

I work across multiple projects, and each one has a different setup: different locations on my MacBook, different conventions, different context. Right now every new session starts blank, so I have to repeat the same instructions again and again for every session.

Codex and ChatGPT already do this with project-level instructions, and this is the single feature I miss most when working in MonoCode. It would be a huge plus.

Steps to reproduce

Current behavior:

1. Right-click a project: the available options today are limited, for example setting a custom logo. There is no field for project instructions or memory
2. Start a new session in that project: it starts with no project context
3. Repeat the same setup instructions manually in every new session

Desired behavior:

1. Right-click a project and edit its own instructions / memory
2. New sessions opened in that project pick those instructions up automatically
3. No repeating the same instructions per session

[hardbeat920]

Hey, wouldn’t creating an AGENTS.md file in the project root through the file explorer already achieve this?

The supported agents should automatically load it when a session starts in that project.

Is this proposal meant to add a more convenient UI for editing AGENTS.md, or does “memory” refer to something beyond static project instructions?

[TheNomadInOrbit]

Thanks, and let me be precise about what I am actually chasing, since my first version was not.

The problem is not that there is no file. It is that instructions only stick when they live in a file. Anything I explain in chat gets summarized away once the session gets long, so I end up explaining the same few project specific things again and again, and the agent quietly stops honoring them mid session with no warning.

There is real evidence for this, not just my impression:

• Codex has an open issue where AGENTS.md rules weaken after auto compaction: Context compaction forgets AGENTS rules: task progress can jump from 97% back to 42% openai/codex#25792
• OpenCode has an open request for a post-compaction reload of AGENTS.md and skills: [FEATURE]: Post-compaction reload for skills, AGENTS.md and instruction files anomalyco/opencode#37627
• Claude Code documents the fix on their side: the project root CLAUDE.md is re-read from disk and re-injected after /compact, and the instructions that disappear are the ones that were only given in chat.

So the property I care about is not just that instructions exist, it is that they survive a long session. That is the part I cannot solve from the UI today.

What I would like, and I am deliberately leaving the approach to you because you know the architecture and I do not:

1. A place per project for a short list of standing instructions, editable from the project menu instead of by hand in the file explorer.
2. Those instructions written to whatever file the active provider actually reads, or one shared file for all of them. I genuinely do not know which of those is the better call across Codex, Claude Code, OpenCode and the rest, so please research that rather than take my guess.
3. Those instructions kept in force as the session goes on, not just at session start, so they do not silently drop out after compaction.

It is only a few lines per project, a handful of rules, and each project is a different case. So yes, it costs a few tokens, and I am fine with that. They are the cheapest tokens in the session if it means I stop repeating myself and the agent stops drifting.

If a shared file plus per provider imports or symlinks is the pragmatic route, that works for me too. I just want to know it is set once, and it holds.