MnemosyneQuick Start
Get Mnemosyne v2.8.0 running in under 5 minutes. This guide covers installation, basic configuration, and your first memory operations — including entity extraction and temporal recall.
Prerequisites
- Python 3.9 or higher
- pip or uv package manager
Install
Install Mnemosyne with semantic search:
# Using pip
pip install mnemosyne-memory[embeddings]
# Using uv (faster)
uv pip install mnemosyne-memory[embeddings]
The [embeddings] extra installs fastembed, which powers semantic vector search. With it, Mnemosyne can:
- Match memories by meaning, not just keywords
- Find "Italian food" when you search "pizza" (semantic similarity)
- Combine vector + text scores for hybrid retrieval
Without [embeddings], you get FTS5 keyword search only.
Your First Memories
from mnemosyne import Mnemosyne
mem = Mnemosyne()
# Store a memory with entity extraction
mem.remember(
"Alice prefers dark mode for the UI. She's the project lead.",
extract_entities=True,
importance=0.8,
)
# Store with LLM fact extraction
mem.remember(
"The deadline was moved to June 15th.",
extract=True,
)
# Temporal recall — find what happened recently
results = mem.recall("UI preferences", top_k=5)
for r in results:
print(f"[{r['score']:.3f}] {r['content']}")
Configurable Retrieval
Mnemosyne v2 lets you tune scoring weights per query:
# Boost text matching over semantic similarity
results = mem.recall(
"exact error code E501",
vec_weight=20.0,
fts_weight=60.0,
importance_weight=20.0,
)
# Emphasize very recent memories
results = mem.recall(
"what did we discuss today?",
temporal_weight=0.6,
temporal_halflife=24.0, # 24-hour half-life
)
Memory Banks
Isolate memories for different projects or contexts:
# Create a bank for work projects
work_mem = Mnemosyne(bank="work")
work_mem.remember("Deployed v2.5.0 to staging.")
# Separate bank for personal
personal_mem = Mnemosyne(bank="personal")
personal_mem.remember("Dentist appointment on Friday.")
Optional: sqlite-vec for Performance
sqlite-vec is an optional performance boost for large datasets (100K+ memories). It moves vector similarity search into native C code inside SQLite. Without it, Mnemosyne uses an in-memory numpy fallback that works fine for most use cases.
pip install sqlite-vecVerify Installation
Confirm everything is working:
python -c "import mnemosyne; print(mnemosyne.__version__)"
# Should print: 2.8.0
What's Next?
- Read the Installation Guide for detailed setup options
- Explore the Python SDK Reference for the full API
- Learn about the BEAM Architecture to understand how memory works
- See Hybrid Search for retrieval tuning details