Memorizz is a Python framework for memory-first agents. It connects persistent memory, models, prompts, and reasoning so an agent can recall earlier context, reuse past experience, and carry what it learns across sessions.
| Memory type | What it holds |
|---|---|
| Episodic | Conversation history, past interactions, and summaries |
| Semantic | Knowledge bases, entity facts, preferences, and personas |
| Procedural | Registered tools, recorded workflows, and reusable skills |
| Short-term | Working context and semantic caching |
| Shared | A shared workspace for agent coordination and handoffs |
Agents are built with MemAgent and MemAgentBuilder in assistant, workflow, or deep-research modes, against cloud or local model providers. Memory is scoped by agent, user, and conversation, and stored on the local filesystem, MongoDB, Oracle, or Notion with a separate vector store, all behind one provider contract.
MetaHarness runs tasks through Codex, Claude Code, OpenHands, DeepSeek, Pi, Hermes, or native MemAgent workers, carrying memory scope across them while handling permissions, approvals, workspace isolation, budgets, and verification. Memorizz also installs as a plugin for Codex and Claude Code, giving both agents a shared episodic memory of past sessions, locally or through a hosted MCP server for a team.
An opt-in continual-learning loop records tool workflows and their outcomes, distills repeated successful procedures into reusable skills, and monitors them, demoting skills that stop working. The project describes itself as experimental, with APIs that may change.
Features
- Memory scoping: memory is isolated per agent, user, and conversation, with retention and forgetting controls
- Knowledge ingestion: files and folders become a knowledge base with configurable chunking, embeddings, and semantic or hybrid search
- Context management: deduplicates retrieved memories, summarizes and compacts history, and exposes recall decisions for inspection
- Storage providers: local filesystem, MongoDB, Oracle, or Notion paired with a separate vector store
- Tools and integrations: Python functions, MCP servers, internet search, governed browser control, sandboxes, and external skills, with a cache for repeatable tool calls
- Multi-agent coordination: delegation to specialist agents, shared memory, and scheduled automations
- MetaHarness: runs a task on an external coding harness or chains several into a plan and compares them side by side
- Coding-agent plugin: session memory shared between Codex and Claude Code, with each session's trajectory viewable in the UI
- Continual learning: captures successful procedures as skills, with shadow evaluation before activation
- Interfaces: Python SDK, interactive CLI, local web UI with run, token, and spend dashboards, and an MCP server mode
