AgentRecall: local-first rule memory for .NET AI coding agents and MCP workflows

[AkbarDizaji]

Hi everyone,

I built AgentRecall, a local-first memory and learning tool for AI coding agents, built as a .NET global tool and NuGet package.

It captures feedback, repeated mistakes, PR review comments, and build/test/lint failures, turns them into structured reusable rules, then ranks and injects the relevant rules back into the agent context through a Claude Code/MCP workflow.

How it differs from built-in memory in tools like Claude:

• It is developer-controlled and project/tool-level, not tied to one chat product
• Memory stays local by default: no cloud sync, no web UI, no API keys
• Rules can be scored, promoted, archived, or superseded instead of stored as plain notes
• Relevant guidance is injected deterministically into context, rather than relying on the model to remember it
• Repeated failures can increase rule confidence and help prevent the same mistake next time

Some current features:

• Must-follow, suggested, and warning rules within a token budget
• Standing rules that apply every turn
• Automatic conflict resolution between rules
• Learning from build/test/lint logs and PR comments
• Seed packs for useful starting guidance
• Project DNA summaries for onboarding agents or developers

I would love feedback from the Microsoft MCP community on:

• Whether this kind of local rule-memory layer fits MCP-based developer workflows
• What examples would make it easier for .NET developers to try
• Whether there are Microsoft Learn or Azure AI agent scenarios where this pattern would be useful

[pdebruin]

Thanks for sharing AgentRecall in the Learn MCP Server Show and tell. Looking through the repo, I couldn't find a direct connection to the Learn MCP Server: it looks like AgentRecall runs as its own MCP server for Claude Code rather than consuming Learn MCP. Could you elaborate on how you see the two fitting together, or point me to where they connect?