Get Mnem

Your AI's memory. Private by design. Ethos · Local-first, model-agnostic
Values · Your memories never leave your device
Purpose · One memory for every AI you use
Where forgetting has a cost, we keep the record.

Get Mnem See how it works

macOS · Windows. Mnem gives Cursor, Claude, Zed, ChatGPT and every MCP client one shared long-term memory, stored entirely on your device. The cloud handles identity, billing, and licensing — nothing else.

One memory, shared by the tools you already use

How it works

Three steps between you and AI tools that remember.

  1. Install the app

    Download Mnem for macOS or Windows. It settles quietly into your menu bar or system tray — with a live memory-field widget showing what it knows.

  2. Connect your tools in one click

    Open the Connections panel and flip on Cursor, Claude Desktop, Zed, and more. Mnem speaks MCP, so anything that supports the protocol just works.

  3. Every AI remembers

    Mention a decision in one app, recall it in another. Preferences, projects, and context follow you across every assistant — and auto-capture only runs where you switch it on.

Your memories never leave your machine

Most memory products are a database in someone else's cloud. Mnem is the opposite.

Local by architecture

Memories, settings, and API tokens live on your device and nowhere else. There is no server-side memory store — we couldn't read your memories if we wanted to.

A tiny cloud, on purpose

Our cloud does exactly three jobs: identity, billing, and license issuance. It stores your email, subscription state, and license and device IDs. That's the whole list.

You decide what's remembered

Auto-capture is opt-in, per app. A never-store topics list keeps sensitive subjects out of memory entirely, and capture modes let you dial recording up or down.

No telemetry by default

The app sends no analytics or usage data out of the box. Routine network traffic is limited to license verification on activation and when you refresh, an update check, a one-time embedding-model download on first run, and — only if you enable capture — distillation calls to the LLM provider you configured.

Read the privacy policy — it's short, because there isn't much we collect.

Built like an instrument, not a database

Live memory-field widget

A living view of what Mnem knows, right in your menu bar or system tray. Watch memories form, strengthen, and connect as you work.

One-click connections

The Connections panel writes the right MCP config for each tool — Cursor, Claude Desktop, Zed, VS Code, and more. No JSON wrangling.

Opt-in auto-capture

Mnem can capture exchanges automatically, but only where you turn it on. Pause or scope it any time — nothing is recorded silently.

Works with every MCP client

Mnem is a standard MCP server. If a tool speaks MCP, it can recall and remember — including tools that don't exist yet.

Never-store topics

List the subjects Mnem must never remember — clients, health, whatever you choose — and they never enter the memory store.

Model-agnostic

Your memory belongs to you, not to a model vendor. Switch apps and models freely; your context comes with you.

Verified performance cold benchmarks, not marketing

Every figure below is read from the output of the open benchmark harness that ships with Mnem — nothing on this plate is typed by hand. The full tables, per-question-type breakdowns, and the vendor claims they sit against are on the benchmarks plate.

†1
LongMemEval end-to-end answer accuracy, all 500 questions — official reader and judge prompts, gpt-4o reader, the benchmark's own judge pinned.
LongMemEval_s · QA ·
†2
Retrieval Recall@5 on the same dataset — the memory layer measured alone, with no answering model and no judge in the loop.
LongMemEval_s · retrieval · questions
†3
Retrieval Recall@10 — the exact context the reader receives. The distance to 100% is the ceiling on the end-to-end number.
LongMemEval_s · retrieval · top-10
†4
On-device recall latency at p95 through the shipped search path. No network round-trip — the column that decides product fit.
Local harness · recall p95 · on-device
† Method notes — what was measured, and how
  1. LongMemEval_s (Wu et al., ICLR 2025), end to end: a gpt-4o reader answers each question from Mnem's top-10 recalled turns, rendered in the benchmark's official session-block history format; every answer is graded yes/no by the official judge prompts with the judge model pinned. Reported over all 500 questions, abstention items included.
  2. Same dataset, retrieval stage only: the share of scorable questions with at least one human-labeled evidence session among the top-5 distinct sessions recalled through the shipped search path, fully on-device. Abstention items excluded per the official retrieval protocol.
  3. Same run at ten recalled sessions — the context window the reader is given, so this bounds what the end-to-end number can reach.
  4. Local recall harness: wall-clock latency through the public search path against a seeded store with 150 interleaved distractors, on-device, no network. p50 and the full table are on the benchmarks plate.

Reproduce any of them: python benchmarks/run_benchmarks.py, python benchmarks/run_longmemeval.py, python benchmarks/run_longmemeval_qa.py — see the benchmarks plate for the exact invocations and the disclosed deviations.

One memory. Every AI. Yours.

Personal plan at $74/month per person — payments open soon, reserve a founding seat today. Business teams — talk to us about an MDM-friendly rollout.

“Memory should be a substrate, not a feature of a model.” — BÄRŌ DYNAMICS, on MNEM
See pricing Mnem for teams