memloom

A local-first memory engine for AI agents. One store on your machine, shared by every tool you use, that persists across sessions and survives switching agents.

npm install -g memloom
memloom init
memloom save "the staging database runs on Postgres"
memloom recall "staging db"

Requires Node 20 or later. No Docker, no database to install: the store is embedded Postgres in a folder at ~/.memloom.

What it does

• One layer across every tool. Claude Code, Claude Desktop, Cursor and your own scripts read and write the same store over MCP.
• Contradictions are kept, not overwritten. A new memory that clashes with an old one keeps both and asks you. Every resolution is reversible.
• Your files are memory too. memloom context add ./notes ingests markdown, text, PDF, web pages and recordings into the same recall, cited back to section, page or timestamp.
• Your agents' history is memory too. memloom import sessions distills months of coding sessions into typed memories.
• Hybrid retrieval. Vector, keyword and entity-graph arms fused by rank in one SQL call, so memloom recall "pg_hba.conf" is exact and memloom recall "that thing about restarting the worker" is semantic.

Works offline with deterministic embeddings. Add OPENROUTER_API_KEY to ~/.memloom/config.env for real embeddings, contradiction detection and entity extraction.

Connect an agent

{
  "mcpServers": {
    "memloom": { "command": "npx", "args": ["-y", "@memloom/mcp"] }
  }
}

Common commands

| Command | Does | | --- | --- | | memloom ui | The viewer: graph, assistant, memories, conflicts, console | | memloom context add <path> | Ingest files or folders as citable source material | | memloom conflicts | List contradictions waiting on you | | memloom reconcile | Review the store, repair what is provably wrong, ask about the rest | | memloom import sessions | Distill agent transcripts into memories | | memloom help | Everything else |

Documentation

docs.memloom.dev

Apache-2.0. Built by Versuno.