On-call memory that
forgets — and proves it.
It remembers your runbooks across every session — and forgets the systems you kill, so you never chase a dead one at 3 a.m.
The real thing builds this from your runbooks — open the app for the full graph.
A real query against the live graph — not a script. For the full forget flip, open the app.
Watch it forget.
Memory you can rely on.
Persistent threads
Every conversation saved — pick it back up later.
Isolated workspaces
Separate knowledge bases, each with its own graph.
Multi-turn context
Follow-ups understand what you just asked.
Runs fully offline
Point it at a local model — nothing leaves your machine.
Verifiable deletion
Hard-delete across graph and vectors — receipt included, re-arm in one click.
Hybrid memory
A knowledge graph and a vector store, working together.
Callable over MCP
Triage and decommission from Claude or Cursor — plus a one-command Claude Code plugin.
Citations that open
Click a chip to read the original runbook — after a forget, it is verifiably gone too.
Memory health, measured
One deterministic staleness score plus a reversible decay loop — free, 0 tokens.
The whole loop — including the verb everyone skips.
Every step is a real call in the codebase — remember, recall, and the one almost no one ships: forget.
- cognee.addRememberingest runbooks
- cognee.cognifyStructuregraph + vectors
- cognee.searchRecallgraph-grounded triage
- cognee.forgetForget the differentiatorverifiable hard-delete
- cognee.pruneResetrebuild from clean
Forgetting, proven.
Most memory products claim they handle stale knowledge. We measure it — decommission three systems, re-ask, and let a blind judge score every answer.
Everyone builds AI that remembers more.
The whole field accumulates — and stale runbooks mislead as often as they help. Lethe makes the opposite move: forgetting the dead thing, on command and verifiably.