DEV Community

Farley Farley (yes, really)
Farley Farley (yes, really)

Posted on

I gave Kiro memory, now it remembers everything...

So here's the thing, I've been using Kiro for about 7 months now, daily, across over 100 projects already and multiple computers. And the single most annoying thing about my experience was it not remembering things I did, or if I happened to be on a different computer which I did them on.

Enter Kiro-Ception, a Kiro Power that gives agent memory across every session, every window, every project, and even across multiple machines. It runs entirely locally (no data leaves your machine by default), indexes everything in the background, and gives Kiro the ability to actually recall past conversations.

The killer feature is that because it's installed as a Power with keyword triggers, you don't have to explicitly ask it to search. You just talk naturally and Kiro automatically pulls up the relevant past context. It feels like it now knows me, and remembers everything we did together, across all projects, the CLI and IDE, and on all my computers.

What it does:

  • Automatically indexes all your Kiro CLI and IDE conversation history in the background
  • Hybrid search, semantic embeddings (understands meaning) + FTS5 keyword (catches exact function names/identifiers)
  • Works across all your projects and workspaces simultaneously
  • Searches are instant (<10ms), cold-starts are <2 seconds from cache
  • Optional multi-node search, find sessions across multiple computers with encrypted transport
  • 100% local by default (CPU-based embeddings, SQLite, numpy). Or plug in Ollama/OpenAI if you want beefier models
  • Auto-triggers on natural language cues ("as we discussed", "what did we do last time", "remind me", "previously", etc.)

Setup is simple, clone the repo, run uv sync and then add the power via the local path in Kiro. It starts indexing immediately, newest conversations first.

Fully open-source, MIT licensed, designed to be privacy/security first, everything stays on your machine. Zero telemetry, zero external calls. The default embedding model downloads once (~80MB) and runs locally on CPU, supports also running higher-end embedding models on your GPU also.

Check it out, star it if you dig it, and let me know if anything's broken. PRs, feedback and contributors welcome.

https://github.com/DevOps-Nirvana/Kiro-Ception

(Yes, the name is a dream-within-a-dream reference. I used Kiro to build a tool for Kiro that remembers everything about Kiro)

Top comments (0)