Multi-agent memory — shared memory for team agents working on the same codebase

[Nagendhra-web]

Feature

When multiple AI agents (Claude Code, Cursor, Copilot) work on the same repo, they should share memory.

Problem: Developer A uses Claude Code, Developer B uses Cursor. Both work on the same repo but their AI agents have separate context.

Proposed solution:

• MEMORY.md is the shared source of truth (already committed to repo)
• Each agent reads and writes to it using the same format
• Agent attribution: "[Claude/JS 2025-04-08]" vs "[Cursor/MR 2025-04-08]"
• Conflict resolution when two agents update simultaneously

Why this matters:
Teams increasingly use different AI coding agents. Memory should work across all of them, not just Claude Code.

[m13v]

MEMORY.md as shared source of truth is the right starting point but conflict resolution is where it gets tricky. git merge conflicts in markdown are manageable when humans resolve them, but when two AI agents write simultaneously you get silent data loss. we found that using structured sections (YAML frontmatter or JSON blocks inside markdown) makes conflict detection more reliable. git can auto-merge non-overlapping keys but will always conflict on adjacent markdown paragraphs. also consider agent attribution at the section level, not line level. 'claude edited auth decisions at [timestamp]' is more useful than per-line tags that clutter the file.

[m13v]

our structured knowledge store with categories, metadata, and conflict-aware schema: https://github.com/m13v/ai-browser-profile/blob/main/ai_browser_profile/db.py

[Nagendhra-web]

Great points. The silent data loss from concurrent AI writes is exactly the hard part. Section-level attribution makes much more sense than per-line tags, and I agree that structured sections (YAML or JSON blocks) would make conflict detection way more reliable than raw markdown paragraphs.

Thanks for sharing the ai-browser-profile schema. The conflict-aware approach with categories and metadata is interesting. I'll dig into it and see what patterns we can bring into the multi-agent memory design.

If you're interested in collaborating on this, PRs are welcome.