agentrecall

AgentRecall is a local-first memory and learning system for AI coding agents, distributed as a .NET global CLI tool and an MCP server. It turns user feedback, build/test/lint failures, and PR review comments into structured, validated rules rather than raw notes. Each rule is parsed into fields such as trigger, rule, do, do_not, reason, and applies_to, then stored in a local database. The tool ranks relevant rules using keyword matching, a deterministic concept graph, scope, and confidence, and buckets results into must-follow, suggested, and warnings within a token budget. It automatically resolves conflicts between rules by scope, supersede, priority, recency, and confidence, and can learn from repeated failures by raising rule confidence or auto-promoting fixes. AgentRecall also compresses memory by detecting and merging duplicate or overlapping rules while keeping an audit trail. It provides deterministic context injection through Claude Code hooks, standing rules that apply to every task, user preferences, project DNA summaries, optional seed packs, and a career-impact pack. A bundled retrieval evaluation reports Precision@1/@3 and Recall@5 and can fail CI on regression. Everything remains on the user's machine: no cloud sync, web UI, or API keys are required.