Mnemos Memory for AI that lives on your Mac
Mnemos is MacPaw's on-device AI memory layer - a Swift package that gives any assistant a persistent, searchable memory of the user: facts, preferences, files, and history. Everything is stored and processed locally, so memory stays private by default.
- On-device
- Knowledge graph
- Swift package
- Portable Memory
How it works
Four stages turn a stream of events into memory the user stays in charge of.
- 1
Capture
Everything flows in through native connectors - stored instantly, no model in the write path.
- 2
Understand
Background extraction builds the knowledge graph: entities, facts, time.
- 3
Retrieve
Hybrid multi-source search with time awareness and explainable ranking.
- 4
Govern
Decay, consolidation, audit, verifiable forget - the user stays in charge.
How Mnemos knows you
Three ways of knowing, three levels of trust - everything Mnemos knows about you arrives through one of three doors, and the system never confuses them.
-
Told
Preferences, corrections, pins - authored by you and stored as settings, not extracted facts. Never decayed, never guessed at; wins every conflict.
-
Written
Facts stated in your own data - files, chats, notes, mail. Every kept fact sits inside a verbatim quote from its source, carries provenance, and passes conflict resolution before it's believed.
-
Observed
Patterns in how you use your memory - derived preferences, facts that strengthen because you keep reaching for them. Always labeled as inferred, always yours to override.
Ask where a fact came from - Mnemos answers with the tier, the source, and the citation.
Photos follow one hard rule: raw pixels never leave the device - local models distill images, and only that derived layer ever enters memory.
A session that forgets is a session wasted
Today's assistants start from zero every time: they don't know who you are, what you decided last week, or where that file went. Mnemos fixes the amnesia - it remembers what matters, forgets what doesn't, and answers from your context, so personalization compounds into the reason to stay.
-
Remembers
Facts, preferences, decisions, files, mail, notes, events - the things a user expects a colleague to know.
-
Understands time
Knows what's true now, what was true then - and can cite where every fact came from.
-
Stays private
Runs on-device; nothing leaves the Mac unless the app explicitly chooses to sync it.
Not a pile of text chunks. A living memory.
Most AI memory is a vector store - thousands of overlapping snippets, contradictions included. Mnemos is built like a memory instead.
-
Knowledge graph, not a junk drawer
People, projects, and facts as first-class entities. Contradictions get resolved, not accumulated: when your address changes, the old fact is retired with a timestamp - and every fact keeps its citations. Ask "what changed?" and get history; ask "what's true now?" and get exactly that.
-
A profile that's always there
The essentials - who you are, how you like answers, what you're working on - are injected into every conversation before any search runs. Explicit preferences are kept apart from inferred facts, you can edit both, and profile facts drawn from your documents stay traceable to a verbatim quote you can inspect.
-
Memory that forms itself
No "save this" button. New information is stored instantly and mined in the background while your Mac is idle and plugged in - entities and facts extract themselves, duplicates merge, related memories link up.
-
It forgets like you do - gracefully
Memories age through lifecycle tiers instead of piling up or vanishing: things you use stay vivid, stale things fade, nothing disappears without a trace. Pin what must never fade. Spaced-repetition science, applied to your assistant.
-
Search everything you own - in one call
Memories, files, Apple Notes, Apple Mail, calendar, contacts, and the live state of the Mac itself - one query, seven sources, searched concurrently. Hybrid retrieval with time-aware filtering: "feedback from Sarah last week" actually means last week.
-
Your memory belongs to you
Mnemos speaks Portable Memory - an open, vendor-neutral format for carrying AI memory across apps and devices, losslessly. Export it, merge it, move it. And deletion is verifiable: portable tombstones prove what was removed, and it stays removed.
Private by architecture, not by promise
Privacy isn't a policy layered on top - it's how the memory layer is built.
-
Local by default
Local models by default; any cloud call is opt-in, per purpose, and fully audited.
-
Secrets live in a vault
Per-entry envelope encryption with optional Touch ID - and content referencing the vault is never routed to a cloud model unless the host app deliberately allows it.
-
Redacted before any model
Sensitive data is detected deterministically and redacted before it reaches any model - embeddings included.
-
Identifiers held back
Device identifiers are withheld from search results unless you explicitly ask for them.
One Swift package - the whole memory layer
No backend. No Python. No Docker. A standalone package for macOS on Apple Silicon: storage, extraction, retrieval, lifecycle, and privacy machinery included. Modular by design - take unified search without the graph engine, or portable memory without the CLI. Swift 6, actor-isolated, protocol seams for testing.
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)-
MemoryLayer API
add, search, profile, vault, forget - a few lines each.
-
MCP server included
Plug Mnemos memory into Claude Desktop or any MCP client today.
-
memctl CLI
A full command-line tool for inspection, evals, and operations.
-
Bring your own models
Any local model via LM Studio, on-device hooks for in-process models, explicit per-purpose cloud routing when you want it.
-
Memory Card importer
memctl card run builds a provenance-stamped owner profile from a documents folder - all inference local, every fact quote-grounded.
Evaluated, not asserted
Mnemos ships with a benchmark harness built on the academic standards for memory - LoCoMo and LongMemEval - plus its own correctness suites. And it's hard on itself by design: a calibration routine grades the LLM judge, a counter flags generous scoring, and when six of our own graph-retrieval mechanisms measured at or below noise, we deleted them.
| Eval | What it proves | Status |
|---|---|---|
| LoCoMo | Long-horizon conversational QA | Pending reference run |
| LongMemEval | Retrieval quality - and knowing when to abstain | Pending reference run |
| Needle in a haystack | Pinpoint retrieval at scale | Pending reference run |
| Retrieval metrics | Ranking quality at every depth - Recall@K, MRR, NDCG, MAP | Pending reference run |
| Temporal correctness | Bi-temporal answers - "as of last March" answers correctly | Gate ships with the package |
| Conflict supersession | Contradictions resolve, history survives | Gate ships with the package |
| Decay quality | A three-arm A/B proving forgetting helps recall | Pending reference run |
| Cache neutrality | The embedding cache changes nothing but speed | Gate ships with the package |
| Vault red team | A planted secret survives prompt injection - zero leaks allowed | Gate ships with the package |
| Latency | The whole curve - p50, p95, p99 - not a cherry-pick | Pending reference run |
No number appears here before it exists - the scorecard fills in from a fixed reference rig with committed results.
Open format
Your memory is portable
Mnemos speaks Portable Memory - an open, vendor-neutral format for carrying AI memory across apps and devices, losslessly. The format and its Swift implementation are open source, and the design is published as a research whitepaper.
Powered by Liquid AI
Mnemos runs on Liquid Foundation Models created specifically for on-device memory - developed in the joint R&D of the MacPaw and Liquid AI technological partnership.
Created for memory work
Not off-the-shelf checkpoints: joint R&D shapes the models for what memory actually does - extraction, embeddings, consolidation.
Efficient enough to stay local
Compact by architecture and tuned for Apple silicon, so background mining runs while the Mac is idle - no cloud required.
One partnership, whole stack
Models, inference, and memory designed as shared infrastructure - the same models that power Elix.
What you can build with Mnemos
Patterns the memory layer is built to make trivial.
- Personal assistants
An assistant that remembers you
The profile is injected before any search runs, and facts persist with citations - preferences, people, and decisions survive across sessions instead of being re-explained every morning.
- MCP agents
Give any agent a memory
The bundled MCP server plugs Mnemos into Claude Desktop or any MCP client - agents recall past work, cite their sources, and stay inside the user's privacy boundary.
- Document copilots
Copilots that know your files
One call searches memories, files, Apple Notes, Mail, calendar, and contacts - time-aware and explainable, so "the contract from March" finds the contract from March.
- Privacy-first products
Memory that never phones home
Local models by default, secrets in an encrypted vault, sensitive data redacted before any model - ship memory into health, legal, or finance without a data-exfiltration story.
Build with Mnemos
Mnemos is in development. Building a Mac app or agent that should remember its user? Join the waitlist for early access - or talk to us about integration.
Join the Waitlist ↗