Harim Choi blog
← All posts
Aug 16, 2026 · 2 min

Monogram

One share action to capture, organize, and search personal knowledge

PKM · Knowledge Automation

Built with Python · EmbeddingGemma · ONNX Runtime · BM25 / RRF · Git Tree API

Monogram architecture showing Telegram, MCP and document inputs through a five-stage validation pipeline into an atomic Git Tree commit and searchable outputs
One path from capture and verification to atomic storage and retrieval

WHY

Useful ideas from news, social media, arXiv, YouTube, and reports were becoming scattered across apps and bookmarks, then disappearing when I needed them again. I wanted one share action to capture any source, preserve why it mattered, organize it automatically, and make it searchable later. The goal was a personal knowledge system that both I and my agents could reuse.

Capture pipeline 5 stages

Orchestrator, classifier, extractor, verifier, writer

Storage boundary 1 Git commit

Related notes and assets written atomically

Agent access 13 MCP tools

The same knowledge can be reused as agent context

HOW

From an unstructured input to a searchable, versioned note

The system separates understanding from writing, then uses Git as both the storage transaction and recovery boundary.

  1. 01 Capture

    Accept Telegram, MCP, Obsidian, or document inputs through the same ingestion boundary.

  2. 02 Structure and verify

    Classify the input, extract fields, and verify the proposed note before writing.

  3. 03 Atomic storage

    Create or update all files with one Git Tree commit so partial writes cannot split the knowledge item.

  4. 04 Hybrid retrieval

    Combine sharded semantic search and BM25 through RRF, with optional graph expansion, MMR, and reranking.

RESULT

Hybrid retrieval diagram showing EmbeddingGemma ONNX embeddings, a sharded INT8 JSONL index, NumPy similarity scan, BM25, RRF, and optional reranking
Custom sharded vector index + BM25/RRF · no FAISS dependency
Monogram dashboard rendered with synthetic example data
Encrypted dashboard · synthetic example data

Contribution

  • Designed the five-stage capture pipeline, atomic Git Tree writer, sharded semantic index, hybrid retrieval path, MCP surface, and recovery controls.
  • Added SSRF defense, secret redaction, cassette evaluation, backup, and a kill switch around automatic capture and retrieval.

Evidence

Source code Capture pipeline, Git storage, retrieval, MCP server, and dashboard