Lethe — long-term memory for AI
A cross-session, cross-model memory engine. What you tell Claude today, GPT remembers tomorrow.
One query, five signals
Keywords, vectors, graph, time and exact keys retrieve in parallel — fused by rank into grounded facts, not vibes.
Memory that stays true
Storing is easy — staying correct is the hard part.
A new statement lowers the weight of the old belief — no stale answers.
The more a fact is repeated or recalled, the more stable it becomes.
A superseded dose or an outdated job title won't resurface as current.
Two realms, one graph
Shared knowledge and private memory live in the same graph — edges may cross, visibility never does.
Designed like a brain
Retrieval, forgetting and consolidation all borrow from biological memory.
Related memories light each other up across the graph.
Unused memories fade; recalled ones grow stronger.
Hot writes settle into cold columnar storage — hippocampus to cortex.
Nothing is silently overwritten; every change is auditable.
One engine, edge to cluster
Embed it as a library, run a single server, or scale out over object storage.
What you can build
One memory, every surface.
It remembers your cat's name.
“Draw my cat” draws your cat.
Scoped by agent and session, never cross-contaminated.
Medical-grade entity linking over curated literature.
Claude Code and Cursor get memory as tools, out of the box.
pip install lethe-client — add, search, done.
Give your AI a memory of you
Start a conversation — from here on, it remembers.