MemoRizz packs five memory systems for agents that forget

What it does MemoRizz is an experimental Python framework for building memory-augmented AI agents. It provides multiple memory systems—episodic, semantic, procedural, short-term, and shared—and lets you back them with Oracle, MongoDB, or a local filesystem. The library also includes agent builders, application mode presets (assistant, workflow, deep_research), and optional add-ons like a local web UI, sandboxed code execution, and scheduled automations.

The interesting bit Instead of manually wiring up vector stores and chat history, you pick an ApplicationMode preset like deep_research and the framework auto-configures the right combination of long-term, short-term, and shared blackboard memory. It also tries to be backend-agnostic for both storage and inference, supporting everything from Oracle AI Database to Apple-Silicon MLX for local LLM inference.

Key highlights

• Five memory types (episodic, semantic, procedural, short-term, shared) with pluggable storage providers.
• Application mode presets (assistant, workflow, deep_research) that auto-enable specific memory stacks.
• Local LLM support via Hugging Face, MLX, Ollama, and OpenAI-compatible servers (llama.cpp, vLLM, LM Studio).
• Optional skills marketplace integration (Vercel Agent Skills, SkillsMP) and sandboxed code execution (E2B, Daytona).
• Multi-tenant memory isolation via user_id and a semantic cache to reduce repeat LLM calls.

Caveats

• Explicitly labeled as experimental/educational software with unstable APIs and no security hardening for production workloads.
• Oracle backend requires additional shell scripts and environment configuration beyond the base install.
• Several advanced features (skills marketplace, WhatsApp automations) are mentioned but only partially documented in the README.

Verdict Worth a look if you are prototyping Python agents and need persistent, multi-user memory without building the plumbing from scratch. Skip it if you need a production-hardened, stable API today, or if you just want a simple vector-store wrapper.