AttogradDB

A lightweight, local-first vector store for semantic retrieval. One SQLite file, no server, no index to maintain. Built for scoped retrieval -- search one project or one session without the rest bleeding in.

Version 1.0.0

Docs: https://gouthamk16.github.io/AttogradDB/ Python API, Claude Code / Cursor / Codex plugins, other MCP harnesses, and marketplace status. Runnable copies: examples/quickstart.py, examples/mcp_memory.py.

Features

• Single-file SQLite storage, no server and no separate index
• Scoped search by project, session, or kind — one value or several at once
• Deletion as a first-class operation -- forgetting matters as much as remembering
• Plaintext, PDF and JSON ingestion with overlap-aware chunking
• Exhaustive exact search: no approximate-recall tradeoff
• Matryoshka dimensions: search at 256-d, keep the full 1024-d on disk
• Project-scoped decision memory over MCP with explicit supersession

Installation

Method 1: Install the PyPI package

pip install attogradDB

Install the optional MCP server with:

pip install "attogradDB[mcp]"

Method 2: Clone and build from source

git clone https://github.com/gouthamk16/AttogradDB.git

Setup and activate python virtual environment

cd AttogradDB
python3 -m venv .venv
source .venv/bin/activate

Install in editable mode with the dev extras and run the tests

pip install -e ".[dev]"
python -m pytest attogradDB/tests

Usage

from attogradDB import VectorStore

store = VectorStore(path="memory.db")          # omit path for an in-memory store

ids = store.add(
    ["retry logic lives in client.py", "we ruled out Redis: no durability"],
    project="payments",
    session="2026-08-18",
)

for doc_id, score, text in store.search("why not redis", top_n=3):
    print(score, text)

store.search("redis", project="payments")      # scoped
store.delete(session="2026-08-18")             # forget
store.close()

add, search, and delete take project / session / kind as a value or a list. Decision memory is a separate table in the same file, exposed over MCP (attograddb-mcp --project-root ...) as remember_decision and recall_decisions. Full docs: https://gouthamk16.github.io/AttogradDB/.

Design notes

Search scans every candidate vector rather than using an ANN index. Measured at 100k chunks of 768 dimensions: 8 ms unfiltered, 3 ms scoped to 2% of the store. An HNSW index is faster unfiltered but roughly 275x slower once a filter is applied, because filtering disconnects its graph while it only shortens an exhaustive scan. Since almost every query here is scoped, and 8 ms is invisible next to an LLM call, the index is not worth its cost. Past roughly a million vectors a scan reaches ~230 ms and that trade changes; CLAUDE.md holds the full numbers.

Roadmap

• Add a method for performance logging.
• LLM based chunking.
• Adding support for more embedding models and indexing methods.
• Adding support for more document types (currently we have pdf, json and txt. Need to add support for docx and images).
• Extract decision candidates from session traces.
• Retrieve long-tail facts using the agent's working context rather than only the user query.
• Verify stored evidence against repository state and compact oversized active decision sets.

Contributing

Contributions are welcome! If you encounter bugs or have feature requests, please open an issue or submit a pull request.

License

This project is licensed under the MIT License. See the LICENSE file for details.