Inside the current prototype
ChatGPT makes a request. Mnemara checks it on your Mac. The memory vault is local; permitted information travels back to the AI.
Demonstrated: ChatGPT + Mac. Phone approvals, cloud storage and broader client compatibility are planned.
Where things live
The Model Context Protocol (MCP) lets AI tools call external services. In this prototype, a paired relay carries requests between ChatGPT and your Mac. It does not host the memory vault.
Your Mac initiates the connection. Approved results return through the relay.
Your Mac
Your Mac must remain online.
The vault stays on your Mac. Approved information doesn’t.
The native app depends on a local owner API/dashboard service. That service must run, but its browser dashboard does not need to be open. The Companion and outbound bridge are separate prototype processes.
Two different decisions
The model uses that result in its active context to answer. It may also decline a tool call or require its own consent.
In the current flow, a locally approved write completes without another ChatGPT “continue” message.
Memory is not the whole conversation
The app or agent around the model assembles what it can use for a response. Mnemara supplies permitted long-term memory; it does not replace the model’s instructions or every project connection.
| Part | Purpose and example |
|---|---|
| Core instructions | How the assistant operates. “Ask before sending anything to another person.” |
| Active context / short-term memory | Information available for this response: the current request, relevant conversation, instructions and any retrieved material. “Tomorrow’s meeting is 30 minutes.” |
| Personal profile / long-term memory | Who you are and how you like to work. “I prefer direct, practical explanations.” |
| Experience and decisions / long-term memory | What happened and what you learned. “Last time the deck had too much text. Lead with the demonstration.” |
| Project knowledge | The detailed work: the current deck, value proposition and financial forecasts in files or an Obsidian vault. |
Personal profile and remembered experience are parts of long-term memory. These categories describe purpose, not exclusive storage locations. A lesson can also exist in a project note.
Retrieval-augmented generation (RAG) means finding relevant information and bringing it into the model’s context to help answer. It can retrieve project documents or long-term memories. It is a technique—not the entire memory framework. Mnemara can use retrieval techniques while also governing storage, revisions and disclosure.
Obsidian can hold lifelong personal knowledge, decisions and project material. Mnemara’s focus is the owner-facing memory and permission experience, not replacing that library. A future connector would need its own scoped access and provenance. An Obsidian integration is not demonstrated or shipping today.
Being honest about the limitations
Shared information reaches the provider. Approved results pass through the relay to ChatGPT. The relay is a transport service, not the vault; that does not make it invisible to the data flow.
Turning off native memory is not a retention guarantee. Providers may retain conversations or process information under their own settings and terms. Mnemara does not erase earlier chats or control information you share directly.
Pause stops future access. It cannot pull back information already disclosed.
Compatibility must be proven. Provider consent, tool behaviour and available connectors can differ between products and clients. Storing memory independently does not automatically make every integration work.
Proof and next steps
Approval-gated corrections through ChatGPT work; that is separate from the unwired native Edit/Delete controls. No independent security certification is claimed.
The bigger picture
That is the direction. The first proof is deliberately narrow: one supported AI connection, a local vault, and an owner who has a say.