AgentKV — The SQLite of Agent Memory

A single-file, embeddable vector + graph store for AI agents. No server required.

pip install agentkv

Quickstart (5 lines)

from agentkv import AgentKV
import numpy as np

db = AgentKV("brain.db", size_mb=50, dim=768)
db.add("Paris is the capital of France", np.random.rand(768).astype(np.float32))
results = db.search(np.random.rand(768).astype(np.float32), k=3)
print(db.get_text(results[0][0]))  # "Paris is the capital of France"

What's New in v0.9.0

• SIMD distance (AVX2): SIMD-accelerated distance computation on supported platforms to improve search throughput.
• Multiple metrics: cosine and l2 both supported and validated.
• Batch insert: add_batch supports bulk inserts (example usage: insert 100 nodes in one call).
• Metadata filtering: Search with where={'color':'red'} filters results by metadata tags.
• Delete / Update: Tombstone + re-insert workflow; deletes decrement the count.
• Count / Iteration: len(db), db.keys(), db.items(), get_all_metadata(offset) supported.
• Windows support: Native Windows wheels produced; CI includes Windows builds.

API Highlights

• AgentKV(path, size_mb=100, dim=1536, metric="cosine") — create/open DB
• add(text, vector, relations=None, metadata=None) -> offset
• add_batch(contents, vectors, metadatas=None) -> List[offsets]
• search(query_vector, k=5, ef_search=50, where=None) — where is metadata filter
• delete(offset), update(old_offset, content, vector, metadata=None)
• len(db), db.keys(), db.items(), get_vector(offset), get_text(offset), get_all_metadata(offset)
• Context helpers: observe(node_id), build_context(node_offsets, max_tokens=2048)

Example (using new features)

import numpy as np
from agentkv import AgentKV

db = AgentKV("demo.db", size_mb=10, dim=128, metric="l2")

# Batch insert 100 random vectors with metadata
texts = [f"item-{i}" for i in range(100)]
vecs = np.random.rand(100, 128).astype(np.float32)
metas = [{"color": "red" if i % 2 == 0 else "blue"} for i in range(100)]
offsets = db.add_batch(texts, vecs, metadatas=metas)

# Search only red items
q = np.random.rand(128).astype(np.float32)
results = db.search(q, k=5, where={"color": "red"})

# Delete + update
db.delete(offsets[0])
new_offset = db.update(offsets[1], "updated text", vecs[1], metadata={"color":"green"})

print(len(db), db.keys(), db.get_all_metadata(new_offset))

Why AgentKV?

Comparison

Installation

pip install agentkv                    # core only
pip install agentkv[ollama]            # + Ollama embeddings/chat
pip install agentkv[all]               # + Ollama + web search

From source:

git clone https://github.com/DarkMatterCompiler/agentkv.git
cd agentkv && pip install -e ".[dev]"

Requires: Python 3.9+, C++20 compiler, CMake 3.15+

Performance

API

from agentkv import AgentKV

db = AgentKV(path, size_mb=100, dim=768)   # Create or open
offset = db.add(text, vector)               # Store + auto-index
results = db.search(query_vec, k=5)         # K-NN search -> [(offset, dist)]
text = db.get_text(offset)                  # Retrieve text
vec = db.get_vector(offset)                 # Zero-copy NumPy view
context = db.observe(offset)                # Predict related context

Examples

See the examples/ directory:

• local_rag.py — Offline RAG with Ollama embeddings
• agent_memory.py — Persistent memory across restarts
• chatbot.py — Interactive CLI with web search

Architecture

Python Agent / LangGraph
        |
  agentkv.AgentKV  (High-level Python API)
        |
  agentkv_core  (nanobind C++ extension — zero-copy, GIL-free)
        |
  C++ Engine
  +-- mmap Storage (single-file persistence, string arena for text + metadata)
  +-- Metadata Store (key/value tags, efficient multi-key filters)
  +-- HNSW Index (vector search) — SIMD-accelerated distance kernels (SSE/AVX2)
  +-- Distance Metrics: Cosine, L2, Inner-Product (selectable)
  +-- Batch Insert / WriteGuard (atomic bulk insert; serialized writers)
  +-- Property Graph (directed edges) + SLB (predictive context)
  +-- Tombstone + Update model (delete=tombstone + re-insert)
  +-- Crash Recovery (CRC checksums, header rollback)
  +-- Thread-safety (GIL released during search; reader-friendly; concurrent writers serialized)
  +-- Platform helpers (`mmap_platform.h`, `simd.h`) — cross-platform (Linux/macOS/Windows)

License

MIT — see LICENSE.

Built with nanobind + scikit-build-core. HNSW algorithm: Malkov & Yashunin (2018).