Open source · MIT License

Claude starts cold.
Retavyn fixes that.

Every Claude session forgets everything — what you worked on, what decisions you made, what you learned. Retavyn gives Claude persistent memory across sessions, automatically.

Retavyn mascot Retavyn brain mascot — persistent memory layer, est. 2026 I remember everything. (please don't test me) est. 2026 retavyn persistent memory layer

The problem

Claude is stateless by design. Every new session, it starts fresh — no knowledge of your project history, your naming conventions, the bug you fixed last Tuesday, or the architectural decision you spent three hours debating.

You end up re-explaining context constantly. Retavyn stores what matters and injects it back automatically at session start. You talk to Claude normally. It remembers.

What it looks like

No commands to learn. Just talk to Claude.

How it works

01
Store memories naturally

Tell Claude what to remember in plain language. Memories are stored in a local PostgreSQL database — on your machine, under your control.

02
Automatic injection at session start

A hook runs at the start of every Claude Code session, dumping all memories into context automatically. No manual steps.

03
Hybrid search when you recall

Full-text search plus semantic similarity via pgvector — recall works whether you use the exact words or just the general concept.

04
Works with claude.ai too

Run in HTTP mode with a Cloudflare tunnel and connect to claude.ai as a remote MCP server. Same memories, any interface.

05
Your memories follow you

Host retavyn on any server and point every machine at it. Work from your laptop, your desktop, claude.ai — they all draw from the same memory pool. One brain, everywhere.

Memory tools

remember

Store a new memory with optional category tagging.

recall

Search memories by keyword, category, or semantic similarity.

update_memory

Edit an existing memory in place.

forget

Delete a memory by ID.

ingest_path

Bulk-import memories from a file or directory.

backfill_embeddings

Generate embeddings for existing memories to enable semantic search.

Stack

Python + FastMCP server

Lightweight MCP server over stdio (Claude Code) and HTTP/SSE (claude.ai). Minimal dependencies, easy to self-host.

PostgreSQL 18 + pgvector storage

Local Docker container. Full-text search via tsvector/tsquery and vector similarity search via pgvector — hybrid retrieval out of the box.

OpenAI / Cohere embeddings

Pluggable embedding providers for semantic search. Supports text-embedding-3-small and embed-english-v3.0.

Claude Code hooks integration

SessionStart hook injects memories automatically. PostToolUse hook refreshes the cache after every write. Zero manual steps.

Cloudflare Tunnel networking

Exposes the MCP server over HTTPS without opening firewall ports. Permanent tunnel with OAuth for claude.ai remote access.

OAuth 2.0 / JWT auth

Custom OAuth flow with JWT tokens secures remote access. Required by the MCP spec for HTTP transport.