Feature

Memory Graph

Entities, relationships, embeddings — one unified graph.

Synapta

How it works

Memory Graph

Synapta sits on ArangoDB and unifies graph, vector, and document storage. Every entity merges by similarity above a 0.85 cosine threshold; every relationship carries its own provenance edge.

Five-tier memory: working (Redis 1h), episodic (90d), semantic (vector + graph, permanent), procedural (triggers), archival (S3, per policy)
Entity resolution merges by similarity — no fragmentation across ingest sources

Primitive

Synapta

Entities, relationships, embeddings — one unified graph.

In practice

Built for replay, audit, and scale.

Five memory layers (Working, Episodic, Semantic, Procedural, Archival) keep retention and latency aligned with access pattern.

Orbit vectors recompute on touched neighborhoods, not the full graph
Open-source store adapters for ArangoDB and Neo4j

# Declare a Reflex once, fire on every match.
reflex:
  name: supplier-risk-spike
  watch:
    cx: entity(supplier).risk_score > 0.75
  fire:
    webhook: https://hooks.example.com/supplier-risk
    cooldown: 1h

Pair it with these primitives.

Packet

Context Packets

The only output shape — provenance assembled in stages.

Reflex

Reflexes

Declarative triggers. CX in, webhooks out.

Citation

Verified Citations

Every fact traces back to its source line.

Try Memory Graph on your stack.

Book a 30-minute walkthrough or wire it up via the API.

Book a demo Back to features