Skip to main content

Personal memory engine for AI agents — zero Docker, SQLite-everything (BM25 + sqlite-vec + SQLite graph)

Project description

kioku-agent-kit-lite

Personal memory engine for AI agents — zero Docker, SQLite-everything.

PyPI Python License: MIT

kioku-agent-kit-lite là phiên bản nhẹ của kioku-agent-kit, thiết kế để chạy hoàn toàn local, không cần Docker, không cần server nào. Mọi thứ đều trong SQLite.

Tính năng

  • Tri-hybrid search — BM25 (FTS5) + Vector (sqlite-vec) + Knowledge Graph (SQLite)
  • Zero Docker — không cần ChromaDB, FalkorDB hay Ollama server
  • FastEmbed ONNX — embedding local, offline-capable (intfloat/multilingual-e5-large)
  • Agent-driven KG — agent tự extract entities → kg-index (không cần built-in LLM)
  • CLIkioku-lite save, search, kg-index, setup, init
  • Python API — import trực tiếp KiokuLiteService vào code
  • Multilingual — tiếng Việt, tiếng Anh và 100+ ngôn ngữ khác
  • 🔜 MCP server — planned (v0.2)

Cài đặt

pip install "kioku-lite[cli]"

Quick Start

CLI

# Lưu memory
kioku-lite save "Hôm nay họp với Hùng về dự án Kioku. Rất productive." --mood work

# Tìm kiếm
kioku-lite search "Hùng làm gì gần đây"

# Index knowledge graph (agent tự extract entities)
kioku-lite kg-index <content_hash> \
  --entities '[{"name":"Hùng","type":"PERSON"},{"name":"Kioku","type":"PROJECT"}]' \
  --relationships '[{"source":"Hùng","rel_type":"WORKS_ON","target":"Kioku"}]'

Python API

from kioku_lite.service import KiokuLiteService

svc = KiokuLiteService()

# Save
result = svc.save_memory("Hôm nay gặp Lan và Minh ở cà phê.", mood="happy")
print(result["content_hash"])

# Search (BM25 + Vector + KG)
results = svc.search("Lan gặp ai hôm nay", limit=5)
for r in results:
    print(r["content"], r["score"])

Agent Workflow

Workflow mẫu để AI agent (Claude Code, OpenClaw,...) sử dụng kioku-lite:

1. Agent lưu memory:
   hash = kioku-lite save "..." --mood work

2. Agent extract entities từ context (dùng LLM riêng của agent):
   entities = [{"name": "Hùng", "type": "PERSON"}, ...]
   rels     = [{"source": "Hùng", "rel_type": "WORKS_ON", "target": "Kioku"}]

3. Agent index KG:
   kioku-lite kg-index <hash> --entities '<json>' --relationships '<json>'

4. Khi cần tìm kiếm:
   kioku-lite search "Hùng làm gì" --limit 5

Thiết kế: kioku-lite không tự gọi LLM — agent chịu trách nhiệm extract entities từ context. Điều này tách biệt hoàn toàn memory store khỏi LLM dependencies.

Cấu hình

Cấu hình qua environment variables với prefix KIOKU_LITE_:

Variable Default Mô tả
KIOKU_LITE_USER_ID default User ID để phân tách dữ liệu
KIOKU_LITE_DATA_DIR ~/.kioku-lite/data Thư mục chứa SQLite DB
KIOKU_LITE_MEMORY_DIR ~/.kioku-lite/memory Thư mục chứa markdown files
KIOKU_LITE_EMBED_PROVIDER fastembed fastembed | ollama | fake
KIOKU_LITE_EMBED_MODEL intfloat/multilingual-e5-large Model name
KIOKU_LITE_EMBED_DIM 1024 Embedding dimensions
KIOKU_LITE_OLLAMA_BASE_URL http://localhost:11434 Ollama URL (nếu dùng Ollama)

Hoặc dùng file .env:

KIOKU_LITE_USER_ID=phuc
KIOKU_LITE_EMBED_PROVIDER=fastembed
KIOKU_LITE_EMBED_MODEL=intfloat/multilingual-e5-large

Benchmark

Benchmark so sánh với kioku-agent-kit (full Docker):

Metric kioku full kioku-lite
Search latency ~2–3s ~1.2s lite nhanh hơn
Precision@3 0.60 0.60 Ngang bằng
Recall@5 1.04 0.89 kit nhỉnh
Infrastructure 3 Docker containers Zero lite

Với cùng embedding model (intfloat/multilingual-e5-large) và cùng Claude Haiku cho KG extraction, kioku-lite đạt chất lượng search bằng kioku full trong khi không cần bất kỳ Docker container nào.

Chi tiết: docs/benchmark.md

Architecture

Xem docs/architecture.md để hiểu thiết kế chi tiết.

Development

git clone https://github.com/phuc-nt/kioku-agent-kit-lite
cd kioku-agent-kit-lite
python -m venv .venv && source .venv/bin/activate
pip install -e ".[cli,dev]"
pytest

License

MIT

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

kioku_lite-0.1.6.tar.gz (49.7 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

kioku_lite-0.1.6-py3-none-any.whl (34.7 kB view details)

Uploaded Python 3

File details

Details for the file kioku_lite-0.1.6.tar.gz.

File metadata

  • Download URL: kioku_lite-0.1.6.tar.gz
  • Upload date:
  • Size: 49.7 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.12

File hashes

Hashes for kioku_lite-0.1.6.tar.gz
Algorithm Hash digest
SHA256 40dab9225d2bd36d21553e44537f8cfc1cd0a251b2c37d40fe3333c0b73b4695
MD5 9c26765def9896269804907287ac7d86
BLAKE2b-256 4753829d5eb51a56e33a166953a92265eb177d795c3a887cf0ffd67f9c4346c7

See more details on using hashes here.

File details

Details for the file kioku_lite-0.1.6-py3-none-any.whl.

File metadata

  • Download URL: kioku_lite-0.1.6-py3-none-any.whl
  • Upload date:
  • Size: 34.7 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.12

File hashes

Hashes for kioku_lite-0.1.6-py3-none-any.whl
Algorithm Hash digest
SHA256 bc5e95557e0636cbee442ee1fdbaace58e703bdd3f08a86534595f6edd141ad9
MD5 a8a981d149c1447488ef95b097888656
BLAKE2b-256 9e037d64702be14a0d04f4600c46e1379f44281b62f197f4189ec3b6944d1c38

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page