lethe-memory 0.4.0

Self-improving memory store for LLM agents: hybrid retrieval, clustered retrieval-induced forgetting, optional LLM enrichment

lethe

Λήθη: the ancient Greek personification of forgetfulness, and one of the five rivers of the underworld.

A memory store for LLM agents that gets better the more you use it. Hybrid BM25 + dense retrieval, cross-encoder reranking, clustered retrieval-induced forgetting (RIF), and optional LLM enrichment at write time.

Most memory tools are static caches - you put strings in, you get strings back by similarity, and the retrieval function never changes. lethe is different: every retrieval teaches it which entries are chronic distractors for which kinds of queries, and it quietly suppresses them over time. No fine-tuning, no extra LLM calls - just bookkeeping inspired by how human memory actually works (Anderson, 1994).

Install and quick start

As a Claude Code plugin (recommended for daily use)

/plugin marketplace add teimurjan/lethe
/plugin install lethe

What happens after install:

• Every session is summarized into .lethe/memory/YYYY-MM-DD.md per project. Plain markdown, edit by hand if you want.
• Claude sees recent memory at session start and calls the recall skill when a past session in this project would help.
• For cross-repo context, the recall-global skill searches every registered project at once (uses lethe search --all under the hood).

Update: uv tool install --upgrade lethe-memory && /reload-plugins

See plugins/claude-code/README.md for the full hook table, config knobs, and debugging.

As a CLI

uv tool install lethe-memory
lethe --version

lethe index                                     # reindex .lethe/memory
lethe search "your query" --top-k 5             # single project
lethe search "your query" --all --top-k 5       # all registered projects
lethe projects list
lethe status

As a Python library

pip install lethe-memory

from lethe import MemoryStore
from sentence_transformers import SentenceTransformer, CrossEncoder

store = MemoryStore(
    "./my_memories",
    bi_encoder=SentenceTransformer("all-MiniLM-L6-v2"),
    cross_encoder=CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2"),
)

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")

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

store.save()
store.close()

Benchmark

Numbers on the full 199,509-turn LongMemEval S corpus, turn-level retrieval, NDCG@10, no leakage. Most memory-tool benchmarks use ~50 sessions at session granularity - a ~2000× easier task. Those 99% numbers don't translate to this setup.

Stage	NDCG@10	notes
Hybrid BM25 + vector (RRF)	0.217	basic retrieval (most popular)
+ cross-encoder reranking	0.293	+35% from semantic reranking
+ clustered+gap RIF (checkpoint 13)	0.312	+6.5% from retrieval-induced forgetting (paired permutation p<0.002, 95% CI excludes zero)
+ LLM enrichment, on covered queries	0.473	+21% on the 75 queries where the answer turn was Haiku-enriched

Scope. The RIF gain is workload-specific. The mechanism targets the chronic-false-positive pattern in a single user's long-term conversation memory. On NFCorpus (a non-conversational medical IR benchmark) it doesn't transfer: three of four variants significantly regress. We diagnose this in the arXiv paper (corpus saturation + workload mismatch) and scope the claim to long-term conversational memory. Use lethe for what it's good at; don't expect it to help on general ad-hoc retrieval.

Full methodology in BENCHMARKS.md. 18 checkpoints (11 failed or null) in RESEARCH_JOURNEY.md. Statistical rigor and the NFCorpus replication in arxiv/paper.tex.

How it works

ARCHITECTURE.md - pipeline diagram, RIF formula, storage layers, entry lifecycle, cross-project search.

RESEARCH_JOURNEY.md - 18 checkpoints from biology-inspired mutation (all failed) through cognitive-science RIF (+6.5% on conversational memory, does not generalize to ad-hoc IR) to LLM enrichment (+21% on covered queries) and statistical rigor with a second-dataset replication.

License

MIT