Mnemos
In developmentMemory for AI that lives on your Mac
Mnemos gives AI assistants lasting memory. It remembers your facts, preferences, files, and history – stored and processed on your Mac by default, with cloud calls only where an app opts in.
How it works
Mnemos turns what happens across your apps into memory your AI can use – while keeping you in control.
-
Remember it as it happens
Conversations, files, and other connected sources are stored right away. No model is needed to write them to memory.
-
Make sense of it
In the background, Mnemos finds people, facts, relationships, and time – and connects them in a knowledge graph.
-
Find the right memory
Mnemos searches across sources, understands when things happened, and ranks results so your AI gets relevant context.
-
Keep memory in your hands
Memories can fade, consolidate, or be removed with a verifiable record. You decide what stays.
How Mnemos knows you
Not every memory comes from the same place. Mnemos keeps three kinds of knowledge separate, so what you tell it doesn't get confused with what it finds or infers.
Told
Your preferences, corrections, and pinned memories. They're stored as settings, never guessed or allowed to fade, and they take priority if memories conflict.
Written
Facts found in your files, chats, and notes. Each fact stays connected to its exact source and is checked for conflicts before Mnemos uses it.
Observed
Patterns in how you use your memory. These are always marked as inferred, and you can correct or override them anytime.
Wonder where a memory came from? Ask. Mnemos can show you its type, source, and citation.
One schema, many doors
- Conversations Available
- Documents Available
- Mail In experiments
- Photos Exploring
Photos follow one hard rule: raw pixels never leave the device – local models distill images, and only that derived layer ever enters memory.
Your AI shouldn't start from zero
Without memory, every conversation is another introduction. Mnemos keeps the context that matters – what you prefer, what you decided, and where things are – so your AI can pick up where you left off.
-
Remembers
Facts, preferences, decisions, files, notes, and events – the context you shouldn't have to explain twice.
-
Understands time
Mnemos knows what's true now and what was true before. Every fact stays connected to its source.
-
Stays local
Memory runs on your Mac. Nothing leaves the device unless the app explicitly chooses to sync it.
Not a pile of text chunks. A living memory.
Mnemos connects what it knows, tracks changes, and brings back the right context when it's useful.
-
People, projects, and facts are connected, not piled into overlapping snippets. When something changes, Mnemos keeps the history and knows which fact is current – with citations attached.
Private by architecture, not by promise
Mnemos is built to keep memory on your Mac and sensitive information away from models that don't need it.
Local by default
Mnemos uses local models by default. Cloud calls are opt-in, purpose-specific, and audited.
Secrets stay in a vault
Sensitive entries can be protected with per-entry encryption and optional Touch ID. Vault content isn't sent to a cloud model unless the app explicitly allows it.
Sensitive data gets redacted first
Mnemos detects and removes sensitive data before it reaches any model – including models used to create embeddings.
Identifiers stay out of search
Device identifiers are withheld from search results unless you explicitly ask for them.
One Swift package. The whole memory layer.
Add memory to a macOS app without building the infrastructure around it. Mnemos includes storage, extraction, retrieval, lifecycle, and privacy – with no backend, Python, or Docker required.
-
Add, search, profile, vault, and forget – the core memory operations take just a few lines of Swift.
let mem = try MemoryLayer(config: cfg)
// Stored instantly – no model in the write path
_ = try await mem.add(RawEvent(sourceType: .chat,
details: "Sarah confirms Q3 slip to May 12"),
extraction: .auto)
// Hybrid, multi-source, time-aware
let hits = try await mem.search("when is launch", limit: 10)Evaluated, not asserted
An honest head-to-head with the leading memory layers – mem0's open-source SDK, Zep (Graphiti), and MemOS – every score shown as measured, across twelve benchmark runs: HippoCamp's three personas, LongMemEval-S and v2, RHELM, LOCOMO, HaluMem, ATM, MemConflict, DynamicMem, and Filegram. Same conversations, same answering model, same judge across systems.
- 89.0%
LongMemEval-S accuracy
Long-horizon memory across sessions, measured on the full 500-question set.
- 100%
Single-session recall
LongMemEval-S single-session user questions – what you tell it, it remembers.
- 93.7%
Conditional conflicts resolved
MemConflict conditional accuracy – context-dependent facts applied in the right context.
- HippoCamp · Adam Accuracy · 123 QA records66.7%42.3% mem069.1% Zep32.5% MemOS
- HippoCamp · Victoria Accuracy · 223 QA records32.3%42.2% mem031.8% Zep36.8% MemOS
- HippoCamp · Bei Accuracy · 235 QA records46.4%31.5% mem045.1% Zep36.2% MemOS
- LongMemEval-S Accuracy · 500 questions · full set89.0%73.2% mem085.2% Zep77.0% MemOS
- RHELM Overall score · 10 personas · 1,305 questions59.1%41.6% mem057.9% Zep46.8% MemOS
- LOCOMO Overall score · 10 dialogues · 1,986 QA pairs64.7%60.3% mem067.5% Zep56.1% MemOS
- HaluMem QA Correct answers · Medium set · 3,467 QA pairs65.3%50.0% mem063.3% Zep55.6% MemOS
- ATM Question score · Main set · 1,013 questions71.5%67.7% mem069.3% Zep60.9% MemOS
- MemConflict Mean accuracy · 30 scenarios · 3 conflict types66.9%58.9% mem060.2% Zep48.4% MemOS
- DynamicMem Mean holistic score · 4 of 10 users · 1,460 rows49.6%41.9% mem039.4% Zep49.1% MemOS
- Filegram Accuracy · Solo set · 20 of 210 personas36.8%34.3% mem036.0% Zep33.0% MemOS
- LongMemEval v2 Accuracy · Small set · 451 questions21.8%16.1% mem022.0% Zep14.9% MemOS
Internal reference runs, September 2026: identical conversations across systems – Mnemos against mem0's self-hosted open-source SDK (not the managed mem0 Platform), Zep's Graphiti engine, and MemOS – retrieval at top-30, answers written by GPT-5, one Gemini judge scoring all sides (Filegram is graded programmatically). Vendor-published scores come from different harnesses, configurations, and judges, and aren't directly comparable with the controlled numbers here. MemConflict and DynamicMem headline scores are means of their category scores. DynamicMem samples 4 of 10 users, Filegram 20 of 210 personas, and LongMemEval v2 is the deliberately harder successor set – no system clears 23% there.
Your memory is portable
Mnemos uses Portable Memory, an open, vendor-neutral format for moving AI memory across apps and devices without losing it. Export it, merge it, or move it when you need to. The format and its Swift implementation are open source, with the design documented in a research whitepaper.
Powered
by Liquid AI
Mnemos runs on Liquid Foundation Models developed through joint R&D by MacPaw and Liquid AI – specifically for on-device memory.
Visit Liquid AI
Built for memory work
These aren't off-the-shelf models. The joint R&D focuses them on the jobs memory actually needs, including extraction, embeddings, and consolidation.
Small enough to stay local
The models are compact by architecture and tuned for Apple silicon, so memory can be processed in the background while your Mac is idle – without requiring the cloud.
One stack, developed together
Models, inference, and memory are developed as shared infrastructure. The same models also power Elix.
What you can build with Mnemos
Add persistent, private memory to assistants, agents, copilots, and other Mac apps.
- Personal assistants
An assistant that remembers you
Preferences, people, and decisions carry across sessions, with their sources attached. No need to explain the same context every time.
- Agents
Give any agent a memory
Wire Mnemos into your agent through the Swift API – it recalls past work, cites its sources, and keeps memory within the user's Mac privacy boundary.
- Document copilots
Copilots that know your files
Search memory, documents, and conversations in one call. Time awareness means "the contract from March" can find the March contract.
- On-device products
Memory that stays local
Local models, an encrypted vault, and pre-model redaction make it possible to build memory into products where sensitive context needs extra care.
Build with Mnemos
Mnemos is in development. Building a Mac app or agent that needs memory? Request a demo, or talk to us about integration.
Request a demo ai@macpaw.comFrequently asked questions
What is Mnemos?
Mnemos is MacPaw's on-device AI memory layer, currently in development. A Swift package gives assistants lasting memory – facts, preferences, files, and history – stored and processed on your Mac by default, with cloud calls only where an app opts in, and built with the aim of privacy by design.
How does Mnemos remember things?
New information is stored the moment it arrives – no model sits in the write path. In the background, while your Mac is idle, Mnemos extracts people, facts, relationships, and time into a knowledge graph, merges duplicates, and connects related memories.
How is Mnemos different from a vector store?
Instead of piles of overlapping text chunks, it keeps a living knowledge graph: people, projects, and facts stay connected, conflicts get resolved, history survives, and every fact keeps a citation. Ask what changed and it knows; ask what is true now and it answers exactly that.
Can I use Mnemos with my own agent?
Yes – Mnemos is a Swift package: add it to your macOS app or agent and call the MemoryLayer API directly – add, search, profile, vault, and forget. Everything stays portable through the open, vendor-neutral Portable Memory format.