Hi Product Hunt Coding agents are smart, but they start every session from zero — re-reading files, forgetting decisions, and burning tokens re-deriving structure. Xanther fixes that with two open-source engines: XCE — Context Engine Indexes your repo into a multi-layer knowledge graph: AST structure (tree-sitter) → LLM summaries → detailed docs → per-module architecture. Your agent queries it over MCP instead of grepping. It also does impact analysis (what breaks if I change this) and traceability (code ↔ architecture). XME — Memory Engine Persistent, cross-session memory: episodic transcripts, an extracted fact graph (decisions, attempts, preferences) with dedup, and a live working-context layer. Your agent stops re-suggesting approaches you already tried and reverted. Both are free and open source (MIT / Apache 2.0), run locally (Neo4j in Docker; XME also has a zero-Docker SQLite mode), and work with any MCP-compatible tool — Claude Code, Cursor, Kiro, Codex, and more. On mini-swe-agent + SWE-bench Verified, adding XCE moved Sonnet 4.0 from 66% → 73.4%, and MiniMax M2.5 + XCE hit 78.2% at ~$0.22/instance (full methodology in the repo). https://github.com/Xanther-Ai/xa... https://github.com/Xanther-Ai/xa... It's early and open — I'd love your feedback on the graph schema, the memory/dedup approach, and the MCP tool surface. Come say hi in Discord: https://discord.com/invite/p27qt... Thanks for checking it out!

Xanther

Context & memory for AI coding agents, over MCP

Context & memory for AI coding agents, over MCP