Documentation

Permanent memory in a few lines.

LeapMemory is not a retrieval wrapper. It is infrastructure for systems that need continuity across months, years, and millions of interactions.

Store a turn

# One call. Persists across four stores.
await lm.turns.ingest(
  user_id="user_482",
  role="user",
  text="Karma is my chihuahua,
        she lives with me in Istanbul.",
)

Recall with provenance

# Multi-signal retrieval. Verbatim + graph + scored.
memories = await lm.recall(
  user_id="user_482",
  query="what kind of dog do i have?",
  k=5,
)
# memories[0].text  -> original sentence
# memories[0].trail -> signals that fired

Walk the entity graph

# Traverse relationships, not just embeddings.
edges = await lm.graph.walk(
  user_id="user_482",
  start="Karma",
  hops=2,
  rels=["IS_A", "LIVES_IN"],
)

Delete a tenant (one command)

# Compliance built into the architecture.
await lm.tenants.delete(user_id="user_482")

# All four stores dropped.
# Physical tenant is gone.
# Nothing recoverable. By design.

Quick start

Four steps to permanent memory.

Getting started

Install the SDK with your package manager.
Create an API key from the dashboard.
Ingest your first turn. Stores fill automatically.
Recall with a query. Get ranked memories back.
Delete a tenant when needed. One physical wipe.
Inspect the trail. Every recall is auditable.

What you do NOT need to do

Chunk your conversations.
Pick an embedding model.
Manage a vector database.
Filter results by tenant id.
Build a graph layer on top.
Worry about TTLs or eviction.

Want full API access?

Hosted access is rolling out via waitlist. Self-host is available now.

Join Waitlist Contact Team