New in v17.22.0 - ✨ Memory

[olimorris]

Hi All,

I'm pleased to share that in v17.22.0, I've added support for memory:

(not the most exciting video, unfortunately)

2025-09-14.00_20_13.-.WezTerm.mp4

Key points:

1. Upon opening the chat buffer, CodeCompanion detects a CLAUDE.md file and adds it as context
2. I leverage the /memory slash command to manually add a CodeCompanion/Chat memory group
3. CodeCompanion detects that the memory group contains references to files and automatically adds them as context

Important

You will need to manually enable memory.

✨ Features

• Automatically add memory files to a chat buffer upon creation, based on memory.opts.chat.default_memory
• Create custom memory groups which are collections of files in your cwd and/or on your disk
• Use parsers to process the content of memory files, extracting file paths that can be included in the chat buffer (see CLAUDE.md imports)
• /memory slash command to choose what memory to add to a chat buffer
• Action Palette item to choose what memory to add to a chat buffer
• .. and much, much more

---

As you're well aware, LLMs don’t retain memory between completions and it can be cumbersome to catch them up in new conversations. Back in January, I tackled this with workspaces. I stand by that idea and the principle. In fact, workspaces are probably the feature I use most in CodeCompanion.

Alas, in their current form, I just feel workspaces being one long JSON file is not conducive to developer happiness.

In #1718, @arnm was kind enough to share his implementation of rules. That is, the rules, files and guidelines that are project specific and useful for sharing with every request to the LLM. With the rise of agents like Claude Code, we have new context files like CLAUDE.md that I expect you're all familiar with.

@arnm's work very much inspired me to think deeply about how CodeCompanion could better support developers who work across multiple projects and as part of large teams.

---

I strongly advise reading about how you can configure memory to get the most out of it. I expect a lot of initial questions can be answered by perusing that page.

Finally, I'm expecting to remove workspaces from the plugin entirely once we get to v18.0.0. But not expecting that to be anytime in the near future.

Thanks and hope you enjoy as much as I do.

Now on to fixing some Claude Code bugs...

Oli

[arnm]

This is awesome work @olimorris !

[olimorris]

Your original gist (here for anyone curious) is still a thing of beauty in my eyes.

[einarpersson]

Glad to see CodeCompanion evolve! I have read through the docs as you suggested but I still couldn't find the answers to these basic questions, perhaps this could be added to the docs?

1. How does on add something to the memory? E.g. Remember that I use the latest version of Svelte 5. To me it looks like this only supports "loading context files". But what about the act of actually adding to the memory? It feels like I miss some piece of the puzzle. Is it editing the files manually currently?
2. What builtin parsers are there? claude and none? And none means "just read the file(s)" as-is?
3. Does using claude require you to use the claude cli, or is it just a "standardized format" that I could use even if I don't use claude cli?
4. It says "Once enabled, the plugin will look to load a common, or default, set of files every time a chat buffer is created. Some of these files are: ..." - Where can I find the complete list? Perhaps add a link to that source file?
5. If I want to add / edit what default files are used, do I need to create a new memory group or can I configure/extend the default memory group?

I will play around with this now! :)

PS. If i'm correct that this feature does not handle "adding to memory", do you have any plans on adding something corresponding to the parsers, eg. a inline slash command that shows a prompt and then calls a "serializer/memorize function" (mirroring the parser) that decides how to add it to memory?

[olimorris]

Great feedback. Have updated the docs

[svalaskevicius]

hi. may I ask where? I came to this pr struggling to find info on pt 1 - does it automatically date memory files? should I just create them as ctx and theyre then static? (if so, how is it "memory" rather than context?)

[olimorris]

https://codecompanion.olimorris.dev/configuration/memory

"Fundamentally, memory is a way of adding persistent context to a chat buffer"

EDIT: This was taken in the spirit of Cursor's rules "Large language models don't retain memory between completions. Rules provide persistent, reusable context at the prompt level."

[olimorris]

To add to part 1, in recent weeks I've added support for a memory tool which was part of my drive to support an LLM's own tooling. It is rigidly based on Anthropic's implementation.

Key point: I've made no attempt to integrate it directly with the memory which is the subject of this post...yet. Reason being, I find in my day to day coding I like to have persistent context (memory) that I share as and when I need, and, saved memories that the tool generates for me if I'm hitting a context limit or wish to save my conversation. I wanted to flesh out that workflow and bounce some ideas around before I potentially integrated the two.

It might well be that this implementation and calling it memory doesn't make sense for a lot of people. I'll be open and say I was super conflicted about calling it memory in the first place. Equally, "rules" (from Cursor's rules) didn't feel like a sensible name either. So actually open to any ideas and better terming of this feature.

[svalaskevicius]

hi. sorry my commented sounded stronger than intended. I was mainly confused as I didn't find it in the docs (might be worthwhile to mention it there - ie what does it NOT do - to avoid confusion). other than that I'm just starting to use the more advanced features and don't really have a good opinion of what is correct :)

thanks for all the work you're putting in here! it's a great plugin :)

[minusfive]

Awesome! I've archived https://github.com/minusfive/codecompanion-agent-rules

[impankratov]

Hi @olimorris! Thanks for this awesome plugin and continuous improvements!
I've tried different combinations of config but there seems no way of dynamically loading custom groups on chat creation.

My use case is that I have a custom group configured, but I want to auto load it (load when code companion chat opens) only for specific folders (by providing a custom function).

Initially though that was possible by conditionally enabling groups but it's only allows for toggling group in memory selection view.

[olimorris]

Hey @impankratov. Could you open up a separate discussion and share a bit more of your use case? I can take a look

[impankratov]

Sure, sorry for the long wait. Opened a new discussion here

[jowu598]

I use workspace extensively in CC to switch between lists of interactive files.
Currently, I use combination of workspace and memory.
In v0.18.0, I noticed that /workspace has been removed, and I’m not sure if there is an alternative solution.
Should we use /memory instead of /workspace?
I configure group, data, and file path in workspace.
Can file path also be configured in /memory?
As far as I remember, AGENTS.md does not contain file path information.