MyMem

Private social memory for AI agents. → mymem.space

Every AI assistant starts from zero about you. Meanwhile the platforms hold years of what you posted, saved, liked, watched and searched for — and they will hand all of it back on request. MyMem turns those exports into one searchable memory your assistant can read, so it stops guessing.

Ask it "what do I actually think about X?" and it answers from your own posts, with the dates and the links. Ask it "who am I?" and it briefs your assistant on what you keep coming back to, the positions you repeat and how you write — all counted from what you actually wrote, never inferred.

Your archive stays yours: it is read where it sits, nothing is uploaded to anyone, and private messages are only ever included if you ask for them.

This repository is the engine behind mymem.space. It is open source (MIT), and the rest of this file is for people who want to run it, read it or add a platform to it.

What works today

Related MCP server: SOMA MCP

Getting your own exports

Request these early — some platforms take 24–48 hours to prepare them. A fictional sample archive ships with this repository, so nothing is blocked while you wait.

The JSON choice matters: pick HTML and the parsers cannot read a thing.

Keep the ZIPs out of the repo and out of screenshots. Then:

npm run mymem -- import path/to/your-export.zip

Running the engine yourself

You do not need to — mymem.space is where MyMem lives. This is for developers who want to see it work end to end.

git clone https://github.com/nixxintools/mymem && cd mymem
npm install
npm test

You should see 84 passing tests and 1 skipped. The skip is expected: it only runs when you ask it to load the real 130MB AI model (MYMEM_TEST_EMBEDDINGS=1).

Then import the sample Instagram archive and search it:

npm run mymem -- import packages/fixtures/samples/sample-instagram.zip
npm run mymem -- search "filter coffee"
npm run mymem -- stats

Expected: 160 items imported, and searches that return dated results with the matched words in [brackets]. Full walkthrough in docs/VERIFY.md.

Two ways of searching

Typing words into a search box only finds what you literally wrote. If you posted "the queue at 7am was down the street" and later search for busy mornings, plain search finds nothing — you never used those words.

So MyMem can search two ways at once:

• By words. Fast, exact, always on. Finds names, handles, phrases you remember verbatim.

• By meaning. A small AI model reads each item and writes down what it is about. Searching then finds things that mean the same thing in different words.

Meaning search is off until you switch it on, because it needs a one-time 130MB model download:

npm run mymem -- embed                              # once, after importing
npm run mymem -- search "morning coffee" --semantic

Results are labelled (words), (meaning) or (words+meaning), so you can always see why something came back.

The model runs on your own CPU. Your posts are never sent anywhere — the only thing that ever crosses the network is the model file itself, downloaded once into data/models and then never again.

If any of that fails — the download, the install, the model — nothing breaks. You get a plain-English warning explaining what went wrong, and search carries on working by words alone. There is no state in which MyMem stops working because the AI part is unhappy.

Let an AI assistant read it

The MCP server hands your archive to Claude (or any MCP client) as twelve tools. Point Claude Desktop at it by adding this to its config file — on Windows, %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "mymem": {
      "command": "node",
      "args": ["C:/path/to/social-memory/apps/server/dist/index.js"],
      "env": { "MYMEM_DB": "C:/path/to/social-memory/data/memory.db" }
    }
  }
}

Restart Claude Desktop and ask it "who am I, according to my archive?" — it will call whoami. For Claude Code, run claude mcp add mymem -- node /path/to/apps/server/dist/index.js.

The server opens the archive read-only: an assistant can ask it anything and cannot change or delete a thing. It answers only from rows in your database, so nothing is sent anywhere.

The twelve tools

The commands

mymem import <export.zip> [--include-dms] [--db <path>]
mymem search "<query>"    [--semantic] [--platform x] [--kind post] [--limit 20]
mymem embed               [--limit n] [--batch 32] [--force]
mymem stats
mymem adapters

npm run mcp                 # start the MCP server (an AI client normally does this for you)
npm run viewer              # browse the archive in a local web page

Direct messages are skipped unless you ask for them with --include-dms. That is deliberate: your DMs are the most sensitive thing in the archive, and importing them should be a decision, not a default.

How it is put together

export ZIP ──▶ parser (one per platform) ──▶ { items, people, profile_facts } ──▶ data/memory.db
                                                                                      │
                                                          word index (FTS5) ──┬── meaning index (optional)
                                                                              │
                                                                        merged results
                                                                              │
                                                                        MCP server ──▶ your AI assistant

Every parser has exactly one job: turn a ZIP into three lists. It never touches the database, the search index, or the AI layer. That is what lets five parsers be written in parallel by five different people without colliding.

What goes in the database

Three tables, frozen early on purpose — changing them changes what every parser has to produce.

• items — one row per thing that happened: a post, reply, comment, DM, like, save, bookmark, video watched, search made, story, or connection. With its date (always UTC), text, media paths, permalink where one exists, and the untouched original record so nothing is ever lost in translation.

• people — everyone who appears anywhere in the archive.

• profile_facts — what the platforms claim to know about you: ad interests, inferred topics, account details.

Plus one optional table, item_vectors, holding what the AI model made of each item. Delete it and everything still works — that is the point of it being separate.

Privacy

• The database, any media, and the AI model all live in data/, which is git-ignored.

• Real export ZIPs are git-ignored too, wherever you put them.

• Direct messages require --include-dms.

• No telemetry, no cloud services, no API keys, no accounts.

• The AI model runs on your machine. Your posts and messages are never sent anywhere. The single network request in the whole project is downloading that model file, once — and you can skip it entirely and still use MyMem.

• Agents developing this project work against the fake sample archives, never against real data.

Contributing

Each parser is one package with one job: turn an export ZIP into items, people and profile_facts. It never touches the database, the search index or the AI layer, which is what lets several be written in parallel without colliding. If you want to add a platform, copy packages/adapter-linkedin (a documented stub), add a sample archive to packages/fixtures, and make the contract test pass — that test is the whole specification.

Next up

See docs/social-memory-ao-build-plan.md. The short version: the parsers for the remaining platforms, then your own exports. Search by words and by meaning both work, and the AI tool layer sits on top of them.