Mnemo:the AI tutor that remembers you, across every platform

Note: All findings are based on self-reported, unverified information from the project description. No external data or traction is available.

The description states that Mnemo is an AI tutor integrated into ChatGPT that remembers a learner’s progress across platforms. It uses GPT-5.6 both during build time (via Codex) and runtime (as the tutoring engine). Mnemo runs on a server deployed via Google Cloud Run, with tools defined using the Model Context Protocol (MCP), and communicates with a React-based widget.

It claims to:

• Maintain a persistent learning path across different AI tools.
• Provide audit trails of how decisions were made in the learning process.
• Distinguish between "proof complete" and "mastered" states.
• Generate live content (quizzes, visualizations) using GPT-5.6.

Inference: The product appears to be a thin client or extension that connects to a user’s existing ChatGPT session and tracks their learning journey through prompts and state management.

The author positions Mnemo as:

• A free, accountable tutoring solution for people who already use ChatGPT.
• An alternative to paid AI tutors that charge an additional subscription on top of the model the learner is already paying for.
• A tool that avoids reinventing AI inference by delegating it entirely to the user's existing LLM.

The evolution of claims:

• Initially, the author frames Mnemo as a personal solution to their own rustiness in coding fundamentals.
• Later, it becomes a broader architectural statement: “One rule: Mnemo runs zero AI inference on its own servers.”
• The positioning shifts from a niche tool for interview prep to a scalable architecture that can serve budget-conscious learners.

Inference: The positioning is rooted in the author’s personal experience and an ideological stance toward minimizing server-side AI usage, rather than any market research or user feedback.

The description states:

• Mnemo targets students, career-switchers, and anyone preparing for interviews on a tight budget.
• These users are assumed to already have access to ChatGPT or similar models.
• The tool aims to avoid pricing out this audience by not requiring an extra subscription.

Inference: The ICP seems to be individuals with basic technical literacy who are using AI tools for learning but lack structured support or tracking mechanisms.

The description states:

• Mnemo is free for users.
• It does not charge for access or usage.
• The architecture avoids running inference on its own servers, which the author sees as a cost-saving strategy.

There is no mention of monetization plans beyond the current free model. No pricing tiers, subscriptions, or revenue streams are described.

Not evidenced: No evidence of any business model beyond "free".

The project was built using:

• GPT-5.6 for both development and runtime.
• Codex for code generation.
• React widget with MCP (Model Context Protocol).
• OAuth 2.1 integration via Cortex (an open-source authorization protocol the author built).
• Google Cloud Run deployment.
• Streamable HTTP, server-sent events, and Zod contracts.

Key technical decisions:

• Zero inference on Mnemo’s own servers.
• Use of MCP tools to interact with ChatGPT.
• Widget state handling via setWidgetState (with a known bug where it overwrites sibling keys).
• Integration with Cortex for cross-platform memory storage.

Inference: The architecture is designed around leveraging existing LLMs rather than building proprietary AI capabilities. This reflects a deliberate choice to minimize infrastructure costs and complexity.

The description states:

• The demo uses real learning data from the author’s own journey.
• It includes a /demo route with no authentication required.
• A roadmap is outlined, including dashboard features, more memory providers, multi-learner studies, and app store submission.

However, there is no evidence of:

• Actual users or customer base.
• Revenue or monetization metrics.
• Usage statistics or adoption data.
• Product-market fit validation.
• Any form of traction beyond the author’s personal experience and prototype.

Not evidenced: No traction signals.

The description does not mention any competitors. It implies Mnemo is unique in its approach to:

• Delegating all AI reasoning to the user's existing LLM.
• Providing cross-platform memory using Cortex.
• Avoiding traditional tutoring model subscriptions.

However, it also notes that most “AI tutors” charge a second subscription on top of the model the learner already pays for — suggesting a competitive landscape exists but is not named.

Inference: Mnemo positions itself as a cost-effective alternative to existing AI tutoring platforms, though no direct competitors are identified.

• Silent bugs and tooling limitations: The author admits to several silent bugs, including datetime schema issues and exact-match module lookups that silently dropped curriculum inserts.
• Dependency on ChatGPT’s behavior: Mnemo relies heavily on ChatGPT's ability to execute tools consistently. The author notes that GPT sometimes drops tool calls after long conversations.
• Limited scalability assumptions: The architecture assumes a single user per session and does not address multi-user scenarios or performance scaling.
• Unproven cross-platform memory: While Cortex is mentioned as the provider, there’s no evidence of successful integration or performance in real-world usage.
• No monetization strategy beyond free access: Without a clear path to revenue, the sustainability of this model is unclear.

Inference: The product is experimental and may not scale reliably due to its dependence on external LLM behavior and limited testing under real-world conditions.

1. How does Mnemo handle situations where the underlying LLM fails to execute a tool call?
2. What specific metrics or feedback have you gathered from early users, if any?
3. Can you demonstrate how the cross-platform memory works in practice? Is it reliable across different AI tools?
4. How do you plan to validate that the tutoring content generated by GPT-5.6 is accurate and pedagogically sound?
5. What are the implications of relying on a single model (GPT-5.6) for both development and runtime, especially in terms of consistency and control?