local-rag-memoria

A tiny, dependency-free, local-only RAG (retrieval-augmented generation) setup: plain markdown files as the source of truth, SQLite FTS5 as the search index, and a local Ollama model to actually answer questions using that context. No cloud API, no vector database, no framework.

Built for a case where all three of those constraints mattered at once: a resource-constrained OpenBSD box where pip chokes on Rust-extension wheels (ruling out most "real" RAG toolkits), no desire to depend on an internet connection or a paid API for something that's meant to keep running indefinitely, and a genuine need to keep the actual facts in a format a human can read and diff in git - not locked inside a database or a vector store's binary format.

How it fits together

fatos/*.md  (source of truth, git-versioned)
     │
     │  sync.py  (rebuild-anytime, disposable index)
     ▼
memoria.db  (SQLite + FTS5 full-text search)
     │
     │  ask.py  (search → build prompt → send to local Ollama)
     ▼
an answer, grounded in whichever fatos/*.md files actually matched

The database is never the source of truth - it's a cache. Delete it and sync.py rebuilds it from the markdown in a few milliseconds. That split matters: it means the actual knowledge stays diffable, greppable, and mergeable in git, while still getting real full-text search instead of "grep and hope."

Usage

# one fact per file, frontmatter for metadata
cat > fatos/some-fact.md <<'EOF'
---
author: alice
date: 2026-01-15
tags: [infra]
---

Whatever you want the agent to remember, in plain prose.
EOF

python3 sync.py                    # rebuild the search index
python3 ask.py "some question"     # search + ask a local Ollama model

$ python3 ask.py "when does the backup run and how long is it kept?"
[1 relevant fact(s) found: example-backup-schedule.md]
The nightly backup runs at 02:20 and keeps 14 daily snapshots plus 6
monthly ones...

Configuration is two environment variables, both optional:

OLLAMA_HOST   default: http://localhost:11434
OLLAMA_MODEL  default: llama3.2

Why not just use <real RAG framework>?

For a real production system with millions of documents, you'd want actual vector search, chunking strategies, reranking, all of it. This isn't that. It's for the much more common smaller case - a few dozen to a few hundred short, hand-written facts - where FTS5's keyword search is entirely sufficient, and the overhead (and dependency surface) of a full vector-DB stack buys you nothing. sync.py is 80 lines; ask.py is about the same. Reading both end to end takes less time than reading a framework's quickstart guide.

License

MIT