Spectron

What is Spectron?

Spectron is an early-preview memory and knowledge layer for AI agents built on SurrealDB. It is designed to store documents, conversations, entities, attributes, relations, embeddings, and traces in a single ACID transaction so agent memory can retain provenance and avoid stitching together separate stores.

The product focuses on typed, provenance-first memory for agent systems. According to the page, Spectron runs on SurrealDB Cloud and sits as a stateless application layer on top of the database, with support for ingesting content, extracting structured data, connecting related facts, and querying memory with traceability.

Key Features

• Provenance on every row: Facts, documents, conversations, and derived memory keep source and trace information, making it possible to inspect where an answer came from.
• Single ACID transaction across memory types: Documents, conversations, entities, relations, embeddings, and traces live in one database transaction, reducing cross-store consistency issues.
• Supersession instead of overwrite: New observations can update or supersede older facts while preserving the record of what changed.
• Typed memory model: The page describes separate memory categories such as authoritative, experiential, reconciliation, elaboration, reflection, consolidation, calibration, and collective memory.
• Retrieval and trace support: The deep-dive references ingest, supersession, hybrid retrieval, and retrieval traces, so reads can be traced back to their supporting memory.
• Built on SurrealDB Cloud: Spectron is presented as running on SurrealDB Cloud, with graph, vector, document, and structured records in one managed substrate.

How to Use Spectron

Users can start by joining the waitlist for invite-only preview access, or evaluate the underlying platform by creating a free SurrealDB Cloud instance. The page also points to an architecture deep-dive and a technical deep-dive for understanding the memory model and seeing the system in action before invite access arrives.

Use Cases

• Agent memory with provenance: Build assistants that need to answer questions with traceable source facts rather than opaque embeddings alone.
• Conversation and document ingestion: Turn chats and documents into structured entities and relations that can be queried later.
• Knowledge reconciliation: Record conflicting or updated statements and track supersession instead of silently replacing older memory.
• Multi-instance or multi-agent memory: Share reconcilable memory across people, agents, and instances while preserving scope and provenance.
• Trust-sensitive retrieval: Use calibration and provenance data to decide when the system should answer, abstain, or surface uncertainty.

FAQ

Is Spectron generally available? No. The page says it is in invite-only preview and that invites roll out in weekly batches from launch week.

Can I evaluate anything before I get an invite? Yes. The page says the substrate underneath ships today, and you can get started free on SurrealDB Cloud or read the architecture and technical deep-dives.

What kinds of data does it handle? The page mentions documents, conversations, entities, attributes, relations, embeddings, and traces.

Does it use separate stores for graph and vector data? No. The page emphasizes one ACID-transactional database rather than stitching together vector, graph, and row stores.

Alternatives

• A vector database plus application-layer memory logic: Common for agent memory prototypes, but the page argues this approach can create seams between stores and make supersession and traceability harder to maintain.
• A graph database with a separate vector index: Useful when relationship traversal matters, but it still requires cross-store coordination for semantic retrieval and provenance tracking.
• A traditional document or knowledge base: Better suited to human browsing and curated content than to typed, transactionally reconciled agent memory.
• Custom memory pipelines built from multiple databases: Flexible, but they shift schema, trust, and consistency work into application code instead of keeping it in the substrate.