[Feature]: add context editing trigger mechanism

[gusye1234]

This is a TODO proposal

Feature Description

Acontext offers tools to edit context, but we never tell devs when to trigger those tools.

Like claude context editing api, it supports trigger mechanism. We should make something like this

Proposed API/Interface

# Python SDK
# Your proposed API here
# TODO

Impact Component/Area

Which part of Acontext would this feature affect?

• Client SDK (Python)
• Client SDK (TypeScript)
• Core Service (Python)
• API Server (Go)
• UI/Dashboard (Next.js)
• CLI Tool
• Documentation
• Other (please specify)

Additional Context

Add any other context, mockups, or examples about the feature request here.

[riturajFi]

Hi @gusye1234 could you assign it to me?

[Pn7Hao]

Hi @riturajFi @gusye1234 , I’ve been working on a similar idea, implemented as a VS Code extension to trigger auto-updates for documentation at the memory level. I hope sharing my approach might be helpful for this feature.

Here is the workflow I designed:

1. Initialization: You need to define the source tree and which file types will be scanned. I use a structure like .memory/ to manage this:
   .memory/
   config/meta.yaml: Defines which files are scanned.
   repo/: Stores the generated docs
   function_a.py.md
   file_a.py.md

2. ChangeDetector: A mechanism to understand the modifications: who changed the file, the timestamp, the diff, or if it is a newly created file.
   3.TrajectoryStrategy: A strategy to traverse and gather relevant context to help the LLM update the documentation accurately.
   4.DefaultSchema: The shape of the document needs to be universal. I recommend structuring it around concepts like skills/.

3. Trigger Rulepack: To avoid modifying docs every time a file is saved, you need specific triggers. For example, manual triggers via the Command Palette or a scheduled Hourly Scan.

Next Steps:
6. I am currently designing a way to feed the documented context from memory back to the agents. I really like your framework and plan to brainstorm how to integrate it later.

I would appreciate any feedback on this approach. I'm definitely looking for more ideas! :>

[gusye1234]

Heey, cool idea.

But this issue focuses on 'simpler' question: when should the editing strategies being used.

Yet, I'm working on the self-learning skill of Acontex right now, which we can discuss many things!

The basic idea is that: we can monitor the agent messages and update the existing skills, or create skills accordingly based on the user feedback and agent experience.

For example, this user is loaded with pptx skill, and he said I want my ppt to always be blue theme, Acontext can locate this and find the SKILL.md of pptx skill, add this requirement. Or Acontext can create a skill for this user call design_ppt_theme with this preferences.

This feature can unify memory and skill in Acontext. I'm still cooking it, but this shares the same value, thus problems with your idea:

• when to trigger update
• how to update the correct files
• how to track the skill changes
• ...

I got a MVP implemented, it's short but got the basic idea: https://github.com/memodb-io/skill-memory.

Look forward to your feedback!