memi

Local-first memory for AI agents. Memories live in one SQLite file on your machine, and agents reach them through an MCP server.

AI agents forget everything when a session ends. memi gives them somewhere to keep the things worth remembering: decisions and why they were made, how you like things done, mistakes not to repeat. The next session can search that instead of asking you the same questions again.

What it does

• Stores memories in a single SQLite file, with no server to run and no account to make

• Finds them by meaning and by keyword, using local embeddings from Ollama or OpenAI

• Keeps global memories (about you, across every project) apart from project memories (about one codebase)

• Lets you mark memories as important, pin the ones that must always be loaded, and sort them into categories

• Runs as an MCP server, so any client that speaks MCP can use it

• Comes with a CLI, a small local web UI for browsing and editing, and agent skills that teach an agent when to save and recall

Related MCP server: engram-mcp

Install

You need:

• Node.js 22 or newer

• One embedder: Ollama (runs locally) or an OpenAI API key

Install it globally to get a memi command:

npm install -g @pinkpixel/memi

Or skip the install and let npx fetch it when it's needed. That's what the MCP setup below does:

npx -y @pinkpixel/memi doctor

The examples below say memi. If you didn't install globally, use npx -y @pinkpixel/memi in its place. To build it from source instead, see Development.

Quick start

The default embedder is Ollama with nomic-embed-text, so pull that first:

ollama pull nomic-embed-text

Check that everything is wired up:

memi doctor

Save a few things, then search for them in different words:

memi add "I prefer tabs over spaces" --global --category preference
memi add "Deploys go through GitHub Actions to Cloudflare" --category fact
memi add "Keep answers short, with no summary at the end" --global --category preference

memi search "indentation style"
memi search "how do we ship releases"

Neither search shares a word with the memory it finds. That's the embeddings at work. (The second memory has no --global, so it belongs to whichever project you ran the command in.)

If you'd rather use OpenAI:

export OPENAI_API_KEY=sk-...
memi config set embedder.provider openai

Keep in mind that memory text gets sent to OpenAI to be embedded when you do this.

Connect it to your agent

Claude Code

claude mcp add memi -s user -e MEMI_AGENT=claude -- npx -y @pinkpixel/memi serve
claude mcp list

-s user makes memi available in every project, which is usually what you want for a memory. MEMI_AGENT is optional (more on that below).

The first start downloads the package, which takes a few seconds. After that npx reuses its cache, so starts are quick, and that cache can hold on to an older version. Use @pinkpixel/memi@latest in the command to always get the newest one. If you installed globally, you can use memi serve as the command instead, which starts fastest.

Other clients

Anything that can launch a stdio MCP server needs a command, its arguments, and optionally some environment variables. In the common mcpServers format:

{
  "mcpServers": {
    "memi": {
      "command": "npx",
      "args": ["-y", "@pinkpixel/memi", "serve"],
      "env": { "MEMI_AGENT": "my-agent" }
    }
  }
}

Restart the client after editing its config. I've only tested this with Claude Code so far, but other clients should work the same way.

What the agent gets

Install the skills

The tools alone don't tell an agent when to use them, so there are three skills in skills/:

They ship inside the package. Copy or symlink the folders into your client's skills directory. For Claude Code that's ~/.claude/skills/, and with a global install it looks like this:

ln -s "$(npm root -g)/@pinkpixel/memi/skills/"* ~/.claude/skills/

Without a global install, grab the skills folder from the repo instead.

The CLI

Known errors print one line starting with memi: and exit with code 1. Commands that need a confirmation won't guess when there's no terminal to ask in. Pass --yes in scripts.

The local UI

memi ui --open

This serves a small memory manager at http://localhost:4747. You can search, filter by scope, category, and importance, add and edit memories, pin them, and delete them. Deleting always asks first.

A few keys help: / jumps to search, n starts a new memory, Esc closes whatever is open, and Ctrl+Enter saves. It works on a phone-sized screen too.

It only listens on your own machine, and it ignores requests that don't come from localhost, since it can delete things. Pick another port with --port, or memi config set ui.port <n>.

Projects, scopes, and the agent name

Every memory is either global or belongs to one project. memi works out the project in this order:

1. A name you pass (--project, or the project argument of a tool)

2. The MEMI_PROJECT environment variable

3. The folder name of the git repo you're in

4. The agent name, stored as agent:<name>

That fourth one is for folders with no git repo. Set it with memi config set agent <name> or MEMI_AGENT. The agent name is also recorded on each memory, so you can see who wrote what.

With none of those, project memories can't be saved, and the error says what to set. Global memories always work.

One thing to know: the git check uses the folder name only. Two repos with the same folder name share a project. Pass an explicit project name if that bites you.

Embeddings

memi config                                   # what's in effect right now
memi config set embedder.provider openai
memi config set embedder.model text-embedding-3-small
memi config set embedder.baseUrl http://localhost:11434

The API key is only ever read from the environment. memi never writes it to a file.

Switching embedders

Different models make vectors of different sizes, so the vectors you already have can't be reused. When you change the embedder, memi notices and turns semantic search off, with a warning. Keyword search keeps working, and nothing is deleted.

To finish the switch:

memi reindex

That drops the vector table, rebuilds it for the new model, and re-embeds every saved memory. Your memories are kept. It asks before it starts. The embedding happens before anything is dropped, so if the new embedder isn't reachable, nothing changes. If you switch back to the original model, you don't need to reindex at all.

If a memory was saved while the embedder was down, it's still saved. It just has no vector yet. memi reindex --missing fills those in without dropping anything.

Configuration

Settings come from ~/.memi/config.json, and environment variables override them.

Setting embedder.provider clears the saved model and URL, so the new provider starts from its own defaults.

Where your data lives

• Database: ~/.memi/memi.db

• Config: ~/.memi/config.json

Set MEMI_HOME to move both. SQLite runs in WAL mode, so you'll see memi.db-wal and memi.db-shm next to the database while memi is running. To back up, copy all three files while nothing is using the database, or run sqlite3 ~/.memi/memi.db ".backup backup.db".

How search works

Searches run two ways and merge the results: a keyword search (SQLite FTS5, with stemming) and a vector search (sqlite-vec, cosine distance). The two ranked lists are combined with reciprocal rank fusion. Importance and recency then give a small nudge, small enough that relevance still wins.

If vectors aren't usable, search falls back to keywords and tells you.

Limitations

• Only tested on Linux so far

• Only tested with Claude Code as the MCP client

• Semantic search always returns its nearest matches, even when none of them are really relevant. There's no relevance cutoff yet

• Search quality depends on the embedding model. nomic-embed-text scores are fairly bunched together, so vague queries can rank oddly

• Two git repos with the same folder name share a project

• The UI can't run memi reindex yet. It shows the command to copy instead

Development

To build from source:

git clone https://github.com/pinkpixel-dev/memi.git
cd memi
npm install
npm run build

Run it with node dist/cli/index.js, or point an MCP client at it with an absolute path:

claude mcp add memi -s user -- node /absolute/path/to/memi/dist/cli/index.js serve

Checks:

npm run typecheck
npm test

The tests are real ones. The end-to-end tests start the actual CLI and MCP server as child processes, and the UI tests run the real frontend modules against the real API. There are no mocks.

The semantic tests need Ollama running with nomic-embed-text pulled, and skip themselves if it isn't. One test, the embedder switch from 768 to 1024 dimensions, also needs a second model. It defaults to qwen3-embedding:0.6b, and you can use another with MEMI_TEST_ALT_MODEL. Without it, that one test is skipped.

The code is split up like this:

src/core/   storage, search, project detection, embeddings plumbing
src/embed/  Ollama and OpenAI providers
src/mcp/    the MCP server and its tools
src/cli/    the command line
src/ui/     the web UI's server
ui/         the web UI's files (plain HTML, CSS, and JavaScript, no build step)
skills/     the agent skills
tests/

License

Apache 2.0. See LICENSE.

Made with 💖 by Pink Pixel