Skip to main content
RhizomeRAG

Four-dimensional memory for AI agents.

In development · Open source · MIT

Most AI memory systems answer one or two questions. RhizomeRAG answers four: what does this mean, how does this connect, when was it true, and how much does it matter. Every retrieval blends all four.

01 / Four dimensions

Why is AI memory more than similarity search?

A flat vector store remembers what sounds alike. That is one signal. Real agents need at least three more.

  • Semantic

    What does this mean?

    Vector embeddings with cosine similarity. Standard, well-understood, and where most RAG systems stop.

  • Structural

    How does this connect?

    A knowledge graph of entities and typed edges. Traversable. Queryable. Lets you ask who, what, and how things relate.

  • Temporal

    When was this true?

    Every fact carries valid_at and invalid_at. The graph supports point-in-time queries so "who worked at X in 2024" returns the right answer.

  • Importance

    How much does this matter?

    PageRank centrality plus access frequency. Recent and well-connected facts rank above forgotten ones, automatically.

02 / Memory types

Four kinds of remembering.

  • Declarative

    Facts, entities, relationships. The stable knowledge an agent accumulates about the world it operates in.

  • Episodic

    Past experiences with state, action, and outcome. Lets agents learn from what they tried before, not just what they were told.

  • Procedural

    Workflows and patterns. Once an agent figures out how to do something well, the procedure store captures the recipe.

  • Temporal

    Validity windows across all of the above. Facts that were true in 2024 do not have to be true in 2026, and the graph knows.

03 / Why open source

The substrate should be shared.

Memory infrastructure for AI agents is a problem bigger than any one product. The companies that try to lock it down end up re-implementing the same primitives behind closed walls. Everyone loses time.

RhizomeRAG is the memory layer behind Hextant. We use it daily and improve it constantly. MIT means you can use it, fork it, embed it, and ship it. If you find a sharper edge to one of the dimensions, send a patch.

04 / Get it

Open source. In early release.

We are stabilizing the public package and repo before the open release. Engaged clients get access to the current build today. If you want to be a design partner for v1, open the conversation.

Hexaxia AI · v2 · 2026RhizomeRAG · Open sourceBuilt by operators