Kindex
The memory layer AI coding agents don't have.
16 followers
The memory layer AI coding agents don't have.
16 followers
A persistent knowledge graph for Claude Code, Codex, Gemini CLI, Antigravity, OpenCode, Cursor, and other AI-assisted workflows. 52 MCP tools, 75 CLI commands, 5 context tiers.
Ito
@barroncaster In that I'm just hoping to help folks and not monetize (MIT licensed), I haven't really thought to track core metrics. Subjectively, I don't feel like I have to "manage" the agent nearly as much. My longest session was 5d 8h with Codex. The whole time, the agent is responsive and helpful and I don't have to remind it about things all the time. When I find myself frustrated, telling the agent to do things differently, asking why it's doing something that way, etc., it's almost always been because Kindex wasn't loaded or being used. It was wildly frustrating to discover a change in Claude Code which had the agent instructed by Anthropic to prefer its own tools and to NEVER TELL THE USER about that instruction.
Token use probably isn't lower. If anything, it's likely higher. But, I use the $200/mo accounts and very rarely hit my usage limits all the same even with multiple agents running simultaneously. The last report from Claude showed that 39% or so of the tokens used were for Kindex. That number seems ridiculously high, but it doesn't (and can't) compare with what token usage would have looked like without Kindex running.
What Codex had to say:
• Yes. It is observational evidence, not a controlled productivity study, but it is concrete.
- Current graph: 6,913 nodes, 18,231 edges, average degree 5.3.
- This repo's committed team graph has 309 durable records in .kin/knowledge.jsonl and 63 commits have updated it.
- A fresh semantic search retrieved the exact cases-runtime cutover decision, its failed first attempt, the required rollback semantics, deployed revision IDs, and the still-open evidence-bucket boundary immediately.
The practical value showed up in the just-completed cutover: the graph preserved that the first attempt had left auth safely dual-authorized, identified the revision-name collision, kept the evidence-bucket task explicitly open, and recorded the exact live proof afterward. That is precisely the information that usually gets lost across agent handoffs, context compaction, deploy logs, and several commits.
The defensible claim is: Kindex materially improves continuity, retrieval, coordination, and evidence discipline for long-running engineering work. It is not yet evidence that it improves velocity by some precise percentage.
I've been using Kindex for months, it has been a complete game-changer. Kindex keeping track of long-running tasks, to-do lists, broad stylistic and architectural goals, as well as permitting multiple agents to use the same shared state while running in parallel -- truly amazing to see.
Kindex has been incredibly helpful when I’m working on long-running features that span multiple repositories.
I can move between codebases without having to rebuild the entire context for my AI agent each time. This saves me from repeatedly explaining the same details and prevents us from wasting time rediscovering work we’ve already completed.
It has made a noticeable difference in my productivity.