I build with AI coding agents all day across a dozen repos, and the thing that kept breaking wasn't capability — it was that every session started from zero. The agent re-learned the same architecture, re-asked the same constraints, re-made the same mistakes. Existing memory tools mostly store transcripts and grep them. That didn't help, because the problem isn't recall — it's structure: what connects to what, what's load-bearing, what's stale. So Kindex is a weighted knowledge graph instead of a log. Typed nodes, edges with provenance that decay over time, and context tiers that adapt to whatever token budget you've got left instead of dumping everything in. My team and I may not be able to tell a huge difference between modern coding agents, but we all know when Kindex isn't loaded. I can have agents collaborate on work across Kindex collab. I can take notes in one instance and act on them in another. Kindex will answer _why_ this line of code looks this way and the .kin knowledge can be shared and edited by your team automatically in the code repo. It's free, MIT, runs local on SQLite, and works with Claude Code, Codex, Cursor, Gemini CLI, and anything MCP-capable. Happy to answer anything.

Kindex

The memory layer AI coding agents don't have.

The memory layer AI coding agents don't have.