John CEO

Company: John CEO

Self-reported basis: The description is entirely self-reported and unverified; it contains no evidence of revenue, customers, funding, or traction.

What the company appears to be: A project that implements a memory system for AI agents with scope-based access control, designed to prevent accidental data leakage by ensuring that facts are only visible within specific conversational contexts (e.g., workspace, channel, thread, DM).

What changed: The project evolved from an earlier public MIT-licensed memory server (qwen-memory-mcp) that used user ID scoping and failed to prevent leaks. This new version introduces a scope-based model where visibility is determined by the conversational path rather than who is asking.

Single most important open question: Is there any evidence of real-world usage or adoption beyond the authors’ own testing?

The description states that John CEO is a standalone Model Context Protocol (MCP) server implementing a memory system for AI agents. It exposes four tools: memory_write, memory_recall, memory_forget, and memory_scopes.

It uses a scope model based on paths like workspace / channel / thread or DM, with the rule that a memory is visible from scope S if and only if its scope is S or an ancestor of S.

The system enforces deterministic isolation by:

• Resolving scopes before the model (deterministic stage)
• Restricting retrieval to visible set only
• Using a ranker (model step) for relevance, conflict resolution, and compression
• Asserting that returned items are in the visible set after the model (hard-fail if not)

The design is said to be implemented using Codex + GPT-5.6, with two key security improvements:

1. Exact-set matching instead of prefix matching to avoid scope leakage.
2. Safe per-segment encoding to prevent path separator forging.

It supports three ranker implementations: a deterministic offline one, GPT-5.6 via OpenAI API, and Qwen via Alibaba Model Studio — all with identical scope guarantees.

Inference: The system is built as a memory primitive that can be consumed by various AI agents (Codex, Claude Desktop, custom bots), but no evidence of actual agent integration or usage exists beyond the authors' own testing.

The description states that this project was built to fix a bug in an earlier version (qwen-memory-mcp) which used user ID scoping and allowed secrets to leak when the same user asked questions in different contexts.

The core claim is:

“A private fact can never surface in a public conversation.”

This represents a shift from asking who is asking to who can hear, positioning the product as a memory primitive with built-in access control rather than a filtering mechanism post-retrieval.

It also claims that:

• The system enforces isolation through deterministic code on both sides of the model.
• It has 47 tests, including adversarial ones.
• It works without cloud infrastructure or Docker (SQLite + sqlite-vec).
• It was built end-to-end in Codex using GPT-5.6.

Inference: The positioning is that this is a secure memory layer for AI agents, not just a tool for storing facts, but one that enforces access control at the point of retrieval and recall.

The description does not name specific customers or target industries.

However, it implies:

• AI agent developers who want to build secure, isolated memory systems.
• Teams using AI agents in collaborative environments, such as workspaces, channels, threads, DMs.
• Organizations concerned with data privacy and compliance, especially those handling sensitive information.

It is not clear whether the product targets:

• End users of AI tools
• Developers building AI agents
• Enterprises or internal teams

Inference: The ICP likely includes developers working on AI agents in environments where access control over memory is critical — but no explicit customer segments are stated.

The description does not contain any information about:

• Revenue streams
• Pricing models
• Monetization strategy
• Subscription plans or usage-based billing

It mentions that the project was submitted to a hackathon and is MIT-licensed, suggesting an open-source approach.

Inference: There is no evidence of a commercial business model. The product appears to be open-sourced (MIT license), with no indication of monetization or pricing structure.

The system is described as:

• Built using Codex + GPT-5.6
• Implemented in Node.js, using MongoDB (though not used in runtime)
• Uses a Model Context Protocol (MCP) server
• Runs with SQLite + sqlite-vec, no Docker or cloud infrastructure
• Supports three ranker implementations: deterministic, GPT-5.6, Qwen
• Has 47 tests, including adversarial ones
• Includes leak assertion and scope resolver stages

It is said to be:

• Deployable in one command
• Fast (npm ci under 4 seconds)
• Designed with security in mind, using deterministic code before and after model steps

Inference: The technical architecture shows a strong focus on security, determinism, and modularity, but there is no evidence of production deployment or scalability testing beyond the authors’ own environment.

The description does not provide:

• Customer base
• Revenue figures
• Usage metrics
• Product adoption
• Market traction

It mentions:

• A live run resolving a contradiction (“we moved reporting from Postgres to Mongo”)
• 47 tests, all green
• Adversarial suite that forces a compromised ranker to return an out-of-scope item and asserts the leak assertion catches it

Inference: The project shows technical maturity in terms of design and testing, but there is no evidence of real-world usage or market traction.

The description does not mention:

• Competitors
• Market landscape
• Alternative solutions

It implies that prior systems (like qwen-memory-mcp) failed to address scope-based access control properly, and this project fixes that gap by modeling audience rather than user identity.

Inference: The competitive context is implied to be in the space of AI agent memory systems, where access control and data leakage are key concerns. However, no direct competitors or market positioning are stated.

• No real-world usage: No evidence of adoption beyond internal testing.
• Limited API exposure: The GPT-5.6 adapter was not exercised against live API due to budget constraints — only Qwen was tested live.
• Hackathon project: Submitted to a hackathon, suggesting early-stage development with no commercial traction or long-term roadmap.
• Open-source only: MIT-licensed, so no clear monetization path.
• No team size or structure details: Only 3 members listed, but no roles or experience breakdown.

Inference: The project is likely in a pre-commercial phase, possibly a prototype or proof-of-concept. It lacks evidence of market validation or product-market fit.

1. What real-world use cases have you tested this system with?
2. Have you built any integrations with actual AI agents (e.g., Codex, Claude Desktop)?
3. How do you plan to monetize this open-source tool?
4. Do you have any customers or partners currently using it?
5. What are the scalability limitations of the current architecture?
6. Are there plans to support more complex scope hierarchies or dynamic scopes?
7. How do you intend to validate that the deterministic gates remain secure under all conditions?