Spectron

Temporal facts are a third key capability, enabling Spectron to track how user preferences, beliefs, and state change over time. Each memory record includes a timestamp and an optional validity period, allowing agents to query only facts that were true at a given moment. This is critical for applications like conversational memory, where an agent must distinguish between current preferences and outdated ones. Spectron also supports state invalidation, where superseded beliefs can be marked invalid in the same ACID transaction. Hybrid retrieval combines vector similarity search with full-text and graph traversal, so agents can find relevant memories by semantic meaning, keyword match, or relationship path—all within one round trip.

Spectron operates on a read-think-write loop executed within a single ACID transaction. In the read phase, the agent issues a SurrealQL query that combines graph traversal, vector search, and temporal filtering to retrieve complete context in one round trip. In the think phase, the agent reasons over the returned result set, which includes documents, relationships, embeddings, and history. In the write phase, the agent persists decisions, updates entities, and triggers events—all within the same transaction boundary. If any part of the write fails, the entire state is rolled back, preventing partial writes. This workflow eliminates the need for compensating logic or two-phase commit between separate systems. Spectron also supports live queries for reactive state and event-driven automation through database triggers.

Concrete use cases for Spectron include agent memory for conversational AI, where an agent retains user preferences and conversation history across sessions. For example, a customer support agent can remember that a user previously tried a specific product and had a certain issue, allowing it to provide informed responses. Another use case is knowledge graph–based context expansion for RAG systems: an agent can retrieve a seed document and then traverse its relationships to find related operational runbooks, improving answer quality. Temporal reasoning is used in applications like inventory management agents, which need to understand what items were available at a past point in time. Spectron also supports hybrid RAG, blending vector and full-text search to retrieve relevant memories even when semantic similarity and keyword occur simultaneously. The outcomes are agents that produce more accurate, context-aware responses and require less glue code.

Spectron targets developers building AI agents, particularly those working on conversational AI, RAG systems, knowledge graph applications, and real-time decision-making platforms. It is built on SurrealDB and leverages the same query language, SurrealQL, which supports multi-model queries across documents, graphs, vectors, and time-series. The technology stack includes built-in authentication, row-level security, live queries, and event triggers, allowing agents to be deployed without separate auth or messaging services. Spectron is available as part of the SurrealDB platform, with plans for cloud and self-hosted deployments. Teams from companies like Samsung, Verizon, Tencent, and PolyAI have adopted SurrealDB for production workloads, demonstrating its scalability. The key takeaway is that Spectron provides a reliable, ACID-compliant agent memory system that eliminates context fragmentation, enabling developers to build agents that remember and reason with trust.