GitHub Repository

A minimal graph engine for grounded AI — records, associates, and retrieves, but never invents. Written in Rust.

12 starsRust

Kremis – Deterministic memory graph for AI agents (Rust)

Name: Kremis – Deterministic memory graph for AI agents (Rust)
Availability: InStock
Author: M2Dr3g0n

by M2Dr3g0n·Feb 17, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainWizardry

Deterministic graph memory vs. embeddings is clever, but fragmentation of AI memory ecosystems is already painful.

Strengths

•Core insight is genuine: replacing probabilistic embeddings with deterministic graph traversal makes agent reasoning fully auditable, not a black box.
•ACID transactions via redb ensure crash safety; strict determinism (no floating-point drift) builds confidence in reproducibility.
•Ships as Rust library, CLI, HTTP API, and MCP server—thoughtful interface coverage for different integration patterns.

Weaknesses

•No evidence of real-world validation; sample data is toy-scale, and integration friction with existing agent frameworks is unclear.
•Graph-based memory is architecturally incompatible with vector-embedding ecosystems (Pinecone, Weaviate); adoption requires wholesale replacement, not composition.

Post Description

Hi HN — I built Kremis, an experimental deterministic memory substrate for AI agents.

I was tired of "black-box" memory where you can't trace why an agent "remembers" or "hallucinates" something. Kremis is an attempt to fix this by using a deterministic graph engine instead of probabilistic embeddings for core state.

Key features:

- Zero Hidden State: Every query result is a concrete path in a graph. You can audit exactly why the AI reached a conclusion. - Strict Determinism: Same input leads to the same graph state. No randomness or floating-point drift in the core logic. - ACID Reliable: Built on redb for crash-safe persistent storage.

How to use it: It ships as a Rust library, a CLI/HTTP API, and an MCP Server. You can plug it directly into Claude Desktop or Cursor to give your AI assistants a verifiable memory.

Development was heavily AI-assisted, and I'm sharing it today to get technical feedback from the Rust and AI community on the architecture.

I'd value your thoughts on: 1. Does a deterministic graph feel like a viable path for long-term agent memory? 2. How can I improve the query ergonomics for complex traversals?

Thanks for any feedback!