Solaris

Solaris is an open-source long-term AI memory system built for developers who need inspectable, honest recall in long-lived AI applications. Released under the MPL-2.0 license and currently in alpha, Solaris treats memory as a three-layer architecture: an append-only archive that serves as the permanent source of truth, a cautiously derived structured memory layer, and an activated set that reflects current belief.

Unlike generic vector-memory wrappers, Solaris separates what happened from what the system remembers. Every derived claim, episode, entity, and relation goes through an editorial review step before becoming a durable memory artifact. This means the system can surface why it remembers something, not just what it remembers. Higher-order patterns are projected as computed views rather than stored as synthetic facts, and divergence between the archive and current belief is made inspectable rather than hidden.

Retrieval is graph-aware: the system indexes entities, relations, and timelines, enabling structured queries that go beyond keyword or embedding similarity. Solaris also ships a shadow-mode deployment option, letting teams run it alongside an existing host system to validate behavior before full adoption. The project includes an MCP server interface, a standalone web API, a detached frontend, a retrieval evaluation harness, and archive-state proof checks.

Solaris is a research-grade tool aimed at AI builders, not a turnkey end-user product. It requires Python 3.12 or higher and optionally Node.js for the web frontend.

Long-term AI agent memory management, auditable AI recall with provenance tracking, graph-structured entity and relation memory for AI systems, shadow-mode memory validation alongside existing AI hosts, editorial review pipelines for AI-derived knowledge, MCP server memory integration, retrieval evaluation and benchmarking for AI memory systems, append-only event archive for AI state history, developer tooling for inspectable AI belief systems