Permanent memory API

Your AI does not need more context. It needs continuity.

LeapMemory is the permanent memory layer for AI applications. Give every one of your users a memory that never expires, grounded in their exact words, physically isolated by architecture.

Prepaid credits. $0.05 per memory saved. Recall is free.
// Traditional memory systems
× Shared vector pools.
× TTL on memories.
× Chunked text retrieval.
× Reconstructed at runtime.
× Filtered by metadata.
// LeapMemory permanent cortex
Verbatim, append-only turns.
Multi-signal retrieval.
Graph walks over entities.
One tenant per memory brain.
Physically isolated by design.
The three claims no other memory API can make

Permanent. Post-RAG. Beyond Vectors.

Three pillars. Each one is the kind of guarantee that requires a different architecture, not a config flag.

01 / PERMANENT

Memories never expire.

No TTL. No summarization. No context-window eviction. Every turn from every conversation is permanently retrievable, exactly as it was spoken.

"Your assistant remembers every user's first conversation, three years from now."
02 / POST-RAG

Turns, not chunks.

RAG was built for documents. Conversation is not a document. LeapMemory stores verbatim turns and fans retrieval across four signals: lexical, semantic, graph traversal, and entity resolution.

"RAG retrieves text. LeapMemory remembers people."
03 / BEYOND VECTORS

Meaning is a graph.

Vectors find similar text. LeapMemory finds connected meaning. Every entity becomes a node. Every relationship becomes an edge. Recall walks the graph, not just embedding space.

"Your user says 'my dog.' We know exactly which one."
Live walkthrough

Watch a memory form.

Pick a fact, ingest it, then ask a question. Every step shows the four stores filling in real time and the retrieval signals that fire.

1Pick a memory
2Four stores, four signals
Canonical
verbatim turns
Graph
entity relationships
Semantic
embeddings + anchors
Resolver
entity resolver
3Ask a question
Retrieved with provenance
Isolation by architecture

One brain per user. Physically.

Every user gets their own databases: their own graph, their own record, their own indexes. Not a tenant_id column. A separate database.

WHY IT MATTERS

Filters leak. Walls do not.

A shared graph with per-row tenant filters breaks the moment one edge of a multi-hop walk misses its filter. In LeapMemory a graph walk cannot cross into another user, because their graph is a different database.

WHAT IT UNLOCKS

Correct multi-hop recall.

Entity resolution never collides across users. "My dog" resolves inside one user's world only. Isolation is not a compliance checkbox here, it is what makes graph memory correct at all.

WHAT YOU CAN SAY

Prove the boundary.

When your enterprise customer asks you to prove their data never touched another customer's data, you have a structural answer. One command deletes a user's entire memory, physically.

The API

Save. Ask. That is the whole surface.

Every user is a tenant with their own memory. Save their turns as they happen, ask when it matters.

Save what a user said
POST /v1/tenants/u_842/turns

{
  "role": "user",
  "speaker_id": "u_842",
  "content": "I moved to Lisbon in May"
}

→ 201 accepted · stored verbatim,
  the memory forms in the background
Ask when it matters
POST /v1/tenants/u_842/recall

{ "query": "where does she live?" }

 {
  "facts": [{ "subject": "u_842",
    "verb": "lives_in", "object": "Lisbon",
    "confidence": 0.97,  }],
  "chunks": [ the exact words, scored ],
  "anchors_found": [  ]
}

Also in there: batch ingest, speaker and time filters, entity anchors. The full reference lives in the docs.

Built for

Systems that cannot afford to forget.

Anywhere identity, context, or cross-conversation continuity matters more than chunked text retrieval.

AI Companions

Permanent identity. Long-term continuity. Year-three recall of a user's first conversation.

Support Copilots

Customers never repeat themselves. Every past conversation is memory, recalled at the moment it matters.

Enterprise AI

Physical tenant isolation. Auditable retrieval boundaries. One-command deletion for compliance.

Memory Platforms

Replace fragile RAG pipelines. Drop in a real semantic infrastructure with provenance.

Under the hood

Four stores. One semantic cortex.

Each store preserves a different aspect of a conversation. Together they form a real memory layer, not a retrieval wrapper.

Canonical

The Words

Verbatim source of truth

Every turn, append-only, exactly as it was said. No chunking. No summarization. No edits. The canonical record of what actually happened.

Graph

The Relationships

Entity knowledge graph

Nodes for people, places, pets, products. Edges for owns, lives-in, prefers, derived-from. Graph walks make recall associative, not just similar.

Semantic

The Meaning

Embeddings + anchor entities

Multilingual embeddings stored alongside the entities each turn anchors to. Lexical and semantic signals fire in the same query. Tenant-scoped by design.

Resolver

The Recognition

Entity resolution cache

The instant lookup that turns a name into the right entity id across languages and time. Hot path for graph traversal. Tenant-isolated keyspace.

Pricing

Pay for what sticks.

Prepaid credits. $0.05 per memory saved. Recall is free, always. No monthly fee, no seats. Per-key spending limits so one runaway agent cannot drain your balance.

Live now

AI is evolving beyond RAG into permanent memory.

LeapMemory gives AI systems a permanent semantic cortex backed by physically isolated infrastructure. The memories persist. The relationships compound. The graph becomes cognition.