Show HN: Kremis – Graph-based memory for AI agents with no hidden state (Rust)

1 point

4 months ago

Hi HN — I built Kremis, a deterministic graph engine designed as a memory substrate for AI agents. Written in Rust, development was heavily AI-assisted.

The core idea: agent memory should be inspectable, deterministic, and honest.

- Same input → same output. No randomness, no floating-point in core. - Every query result traces back to a concrete graph path — no hidden state. - Zero pre-loaded knowledge. All structure emerges from ingested signals. - ACID transactions via redb. Crash-safe persistent storage.

It ships as a library (kremis-core, pure Rust, no async), an HTTP API + CLI, and an MCP server so AI assistants like Claude can query the graph directly.

Current state: v0.3.1, experimental, ~277 tests, CI on 3 OS, Docker image.

I'd value feedback on: - Does the deterministic graph approach make sense for agent memory? - API ergonomics — is the query model (lookup/traverse/path/intersect) intuitive? - What failure modes should I prioritize testing?