lethe-memory 0.12.0

Persistent memory store for LLM agents. Python bindings to the Rust lethe core.

lethe-memory

Python bindings to the Rust core of lethe, a self-improving memory store for LLM agents. Built with PyO3 + maturin, distributed as wheels.

Install

pip install lethe-memory

Use

from lethe_memory import MemoryStore

store = MemoryStore("./my_memories")  # encoders default to MiniLM

store.add("I prefer window seats on flights",      session_id="trip")
store.add("My wife needs aisle seats",             session_id="trip")
store.add("I work at Google as a software engineer", session_id="work")

for hit in store.retrieve("What are my travel preferences?", k=5):
    print(f"  [{hit.score:.1f}] {hit.content}")

store.save()

The wheel ships the Rust lethe-core retrieval pipeline (BM25 + dense hybrid, cross-encoder rerank, clustered RIF), the same code path the CLI uses, exposed via PyO3 with Python::allow_threads around all native work so multi-threaded Python callers see real parallelism.

Constructor knobs

MemoryStore(
    path,
    bi_encoder="Xenova/all-MiniLM-L6-v2",
    cross_encoder="Xenova/ms-marco-MiniLM-L-6-v2",
    rif_config=None,                # RIFConfig(n_clusters=30, ...)
    k_shallow=30,
    k_deep=100,
    confidence_threshold=4.0,
    dedup_threshold=0.95,
)

See also

• Project landing page: architecture, benchmarks, research journey
• lethe-cli (Homebrew: lethe): the standalone CLI

License: MIT.