Kiroku automatically extracts knowledge from every Claude conversation, embeds it in a local vector database, and makes it searchable via MCP. Your AI assistant finally remembers.
Three invisible layers that give Claude Code a memory, without changing your workflow.
Routes Claude API traffic through a local proxy that side-records every conversation turn to a JSONL queue. Zero latency impact.
A background worker processes each turn with an LLM to extract structured entities, facts, and relationships. Fully async.
Facts are embedded with bge-m3 and stored in SQLite + sqlite-vec. Claude queries them automatically through MCP tools.
A half-life system keeps frequently-accessed memories warm and lets stale knowledge fade. Your memory stays relevant.
Three independent modules. All data stays local.
Everything you need for AI-native project memory.
Vector similarity + keyword matching. Claude finds relevant memories automatically through MCP, no manual tagging needed.
AWS keys, API tokens, GitHub PATs — all automatically detected and redacted before storage. Your secrets stay secret.
Personal preferences and global knowledge persist across all projects. Claude remembers "always use bun" everywhere.
Half-life decay keeps memories fresh. Frequently accessed facts stay hot, unused ones cool down naturally.
All data stored in a local SQLite database. No cloud sync, no third-party storage. Your knowledge, your machine.
Claude can run read-only SQL queries against your memory database for complex analysis, timeline queries, and cross-project stats.
Start free. Upgrade when your memory grows.
No. The proxy records conversation turns locally. Extraction uses your own LLM API key (OpenRouter, Ollama, etc.). All extracted knowledge stays in your local SQLite database. The only network call Kiroku makes is the extraction LLM request.
Any OpenRouter model (Gemini Flash, Claude Haiku, Llama, etc.) or local Ollama models. The default is Gemini 2.0 Flash via OpenRouter — fast, cheap, and high quality.
Yes. Kiroku transparently proxies to Anthropic's API. It works with any Claude plan that supports the API, including Pro and Max.
Free tier limits you to 500 stored facts, 50 daily extractions, and uses a basic extraction prompt without vector search. Pro removes all limits, enables vector similarity search with bge-m3 embeddings, and gives you the premium extraction prompt with higher accuracy.
Yes. Kiroku is designed to be fully self-hostable. Free tier works completely offline. Pro features require a license key for the premium extraction prompt, but all data processing happens locally.
After purchase, run kiroku activate <your-license-key>. The key is tied to your machine ID. You can deactivate and transfer to another machine anytime with kiroku deactivate.
Install in 30 seconds. No sign-up required for the free tier.