Memryzed
Local-first cross-agent memory for AI coding agents.
Every AI coding agent forgets everything between sessions, and none of them share anything with each other. What you tell Cursor on Monday is gone by Tuesday, and Claude Code never knew about it at all.
Memryzed fixes both halves: memory that persists across sessions, and memory that's shared across agents. Tell one agent something, and another agent can recall it.
How it works
- Runs locally over stdio as an MCP server. Works with Claude Code, Cursor, Kiro, Copilot CLI, Codex, and Continue.
- Hybrid retrieval: vector + full-text + lexical, fused with reciprocal rank fusion.
- A bundled BGE-small embedder runs locally via ONNX. No API keys, no outbound calls.
- All data lives in one local SQLite file. No accounts, no telemetry, no cloud.
- A background engine captures your agent conversations and proposes facts; you approve, edit, or reject them from a CLI review queue, so nothing is stored without you.
Three layers of memory
- Global — preferences and facts that follow you across every project and machine.
- Project — repo-scoped facts: build commands, conventions, ownership, decisions.
- Session — per-task working state so you can resume exactly where you left off.
Tools (9)
recall, remember, forget, list_memories, extract_from, checkpoint, resume, list_sessions, end_session
Install
macOS, Linux, WSL:
curl -fsSL https://memryzed.com/install.sh | bash
Windows PowerShell:
irm https://memryzed.com/install.ps1 | iex
Then:
memryzed init # one-time setup
memryzed install # auto-wire into detected MCP clients
memryzed doctor # verify everything is working
Quality
On LongMemEval-S (a long-conversation recall benchmark), un-reranked and untuned: R@1 0.587 / R@5 0.842 / R@10 0.902. The result data and methodology are in the repo.
Privacy
Local-first by default. No accounts, no telemetry, no network calls. Everything is in one SQLite file on your machine that you can back up or delete.
Apache-2.0 · github.com/memryzed/memryzed · memryzed.com