memme 0.1.0

Edge-first AI memory engine powered by DuckDB

MemMe — Python SDK

Edge-first AI memory engine for Python. Store, search, and manage long-term memories for AI agents and applications. Powered by an embedded DuckDB database with built-in vector search, BM25 full-text search, knowledge graph, and forgetting curve — all in a single file, no external services required.

Installation

pip install memme

Quick Start

Basic Usage

from memme import MemoryStore

# Local ONNX embeddings (no API key needed)
store = MemoryStore("memory.duckdb")

# Add memories
store.add("User prefers dark mode", user_id="alice")
store.add("User drinks coffee every morning", user_id="alice")

# Search memories
results = store.search("morning routine", user_id="alice", limit=5)
for r in results:
    print(r["content"], r["score"])

With OpenAI-compatible Embeddings

store = MemoryStore(
    "memory.duckdb",
    embedder="openai",
    api_key="sk-xxx",
    base_url="https://api.openai.com/v1",
    embed_model="text-embedding-3-small",
    dims=1536,
)

LLM-powered Smart Extraction

Automatically extract structured facts from natural language conversations:

store = MemoryStore(
    "memory.duckdb",
    embedder="openai",
    api_key="sk-xxx",
    llm_api_key="sk-xxx",
    llm_model="gpt-4o-mini",
)

# Extracts facts: location=Tokyo, activity=learning Japanese
store.add_smart(
    "I moved to Tokyo last year and started learning Japanese",
    user_id="alice",
)

results = store.search("where does alice live", user_id="alice")
# -> "User moved to Tokyo"

Session and Episode

Organize memories into sessions (conversations) and episodes (meaningful segments):

# Create a session
session = store.create_session(user_id="alice", metadata={"topic": "onboarding"})

# Add messages to the session
store.add_message(session.id, role="user", content="Hi, I'm new here!")
store.add_message(session.id, role="assistant", content="Welcome! Let me help you get started.")
store.add_message(session.id, role="user", content="I prefer dark mode and minimal notifications.")

# Compact session into long-term memories
store.compact_session(session.id)

# Later, search across all memories
results = store.search("user preferences", user_id="alice")

Features

• Edge-first — Runs locally with embedded DuckDB, no external services required
• Single-file storage — Vectors, knowledge graph, FTS index, and history in one .duckdb file
• Multi-channel retrieval — Vector + BM25 + Entity Graph + Temporal (RRF fusion)
• LLM-powered extraction — Automatic fact extraction with temporal date resolution
• Forgetting curve — Ebbinghaus-inspired memory decay with stability reinforcement
• Privacy controls — LocalOnly, Syncable, EncryptedSync per-memory levels
• Cross-platform — Python, Node.js, Rust, Swift, WASM

Benchmark (LoCoMo)

Category	MemMe	mem0	mem0-graph
Single-hop	80.50	67.13	65.71
Multi-hop	55.76	51.15	47.19
Temporal	59.38	55.51	58.13
Open-domain	74.55	72.93	75.71

Links

• GitHub Repository
• Documentation
• Rust Crate (memme-core)
• npm Package (memme)

License

Apache-2.0. See LICENSE for details.