mem0-nix

Nix packaging wrapper for upstream mem0ai/mem0 — the self-hosted memory layer for AI agents.

This repo tracks upstream mem0 HEAD and exposes two Nix packages consumed by the nixos-config flake:

Package	What it builds
mem0-server	Python app (FastAPI + uvicorn) from server/ — REST API + MCP endpoint
mem0-dashboard	Next.js standalone server from server/dashboard/ — proxied by nginx in the NixOS module

Usage

As a flake input (in nixos-config)

inputs.mem0-nix.url = "git+ssh://git@git.nimmog.uk:2222/nimmo/mem0-nix.git";
inputs.mem0-nix.inputs.nixpkgs.follows = "nixpkgs";

mem0-server = inputs.mem0-nix.packages.${system}.mem0-server;
mem0-dashboard = inputs.mem0-nix.packages.${system}.mem0-dashboard;

Build locally

nix build .#mem0-server
nix build .#mem0-dashboard

Update upstream pin

./update.sh

This bumps the pinned commit in flake.nix and attempts a fetch to discover the correct SHA. After the first build, replace fakeSha256 / fakeHash with the computed values.

Architecture

The wrapper mirrors the librarry-nix pattern: this repo only packages the upstream source. The NixOS module (modules/services/mem0.nix in nixos-config) handles service wiring — Postgres/pgvector, systemd, nginx, SOPS secrets.

Current packaging patches in packages/mem0-server-env-config.patch do two things:

• allow selecting Ollama-backed LLM/embedder providers via environment variables, matching the NixOS service module wiring
• turn infer=true writes that produce no storable memories into a clear API failure instead of a silent 200 with empty results

Upstream structure

• mem0/ — core Python library (built as mem0ai via hatchling)
• server/ — FastAPI server with Alembic migrations, auth, MCP routers
• server/dashboard/ — Next.js dashboard UI

No Docker

The entire stack runs as native NixOS services. Postgres + pgvector are provided by services.postgresql in the NixOS module, not a container.