Find a hotel in Lisbon.Current request + bounded recent context
Open source · macOS · solo maintained
An AI assistant
trying a different way to remember.
When enabled, Moki uses Jev to judge which topics and entities matter to your request, then retrieves relevant records from local memory instead of putting one fixed memory file into every prompt.
Developer ID-signed and notarized. No Moki account required.
Memory recalledBacked by 2 messages
Stored locallySQLite on your Mac
The point
Useful memory should not have to fit in two prompt files.
A fixed USER.md or MEMORY.md is simple, but its whole contents consume context on every turn. Moki is an experiment in routing a larger structured memory collection and loading only a small relevant set.
The bet is not that Moki remembers more today. The bet is that typed descriptor routing can make a larger memory system practical without sending the raw archive to the router.
Jev-routed memory
Route through topics.
Retrieve locally.
Jev scores compact topic and entity descriptors against the current request. Moki then retrieves candidate memories from SQLite, follows bounded graph relationships, validates their evidence, and adds at most eight complete records to the prompt.
Score descriptors
Lisbon triprequired
Alexrelevant
Topics and entities, not raw memory textValidate and expand
SQLite eligibility
+ one graph hop
+ lexical route
Relevant records only
“Alex requires step-free access.”
Up to 8 whole records · 8,000 charsDescriptors, not the archive
Jev sees bounded topic and entity labels. The raw memory records remain in local SQLite during routing.
Evidence still matters
Selected records must have current support. Editing or deleting a source makes its old evidence stale.
An explicit experiment
The routing is shipped and bounded, but live precision, recall, and latency are not established yet.
The rest of Moki
Bring your own brain.
Choose its hands.
Models
Use model access you already have.
Add a DeepSeek API key or sign in with ChatGPT for supported Codex models. Switch models per conversation.
DDeepSeekAPI keyConnected
◌ChatGPTCodex sign-inConnected
Connect through MCP.
Add local or remote MCP servers. Enable individual tools instead of handing everything to every conversation.
⌕↗⌘+
Talk, type, or capture.
Use native push-to-talk dictation or ask about a region of your screen when the selected model accepts images.
Control
Optional means optional.
Memory recall, automatic learning, smart tool loading, and Jev routing have separate settings. Moki keeps working with local fallbacks when optional context services are disabled.
Memory recall
Automatic learning
Smart tool loading
Open and yours
No Moki account. No telemetry. Local storage.
Moki stores its own data on your computer. Credentials are encrypted with macOS-backed Electron storage. Your configured providers and tools still receive the requests required to work, and Moki tells you where those boundaries are.
- ✓ Apache 2.0 source code
- ✓ Local conversations and memory
- ✓ No required Moki account
- ✓ No product telemetry
A useful boundary
Moki is not for everyone.
It is an early, solo-maintained desktop assistant for people who want to configure and inspect their setup.
Probably for you if
- You use a Mac with Apple Silicon
- You want control over providers and tools
- You want to experiment with routed long-term memory
- You are comfortable with early software
Probably not for you if
- You need a completely offline assistant
- You want bundled local models or cloud sync
- You need Windows, Linux, or Intel Mac support
- You require enterprise support guarantees
Moki 0.1.0
Make your assistant
a little more yours.
Available now for Apple Silicon Macs.
Signed and notarized · Approx. 155 MB · macOS