Mnemex: Anchored Decision Memory for Coding Agents

What the company appears to be: Mnemex is a local, deterministic system for managing code decision memory in AI agents. It uses content hashing to anchor decisions and prevent contradictions while allowing legitimate refactors. The system operates fully offline with no telemetry or cloud dependencies.

What changed: The project is presented as a hackathon submission (OpenAI 2026) that implements a guardrail system for coding agents. It builds on the idea of "anchored decision memory" where decisions are tied to code content hashes rather than abstract rules.

The single most important open question: Does Mnemex actually solve the core problem it claims — distinguishing between real contradictions and legitimate evolution in AI agent code changes — or is this a theoretical framework that hasn't been tested at scale?

Analysis basis: This analysis is based entirely on the self-reported project description provided by the author. No external verification, traction data, revenue figures, or customer information are available.

The description states Mnemex is "a local SQLite brain" that:

• Uses SQLite + FTS5 for keyword retrieval (BM25)
• Parses Python and TypeScript/TSX code
• Extracts symbols, calls, and imports
• Content-hashes every symbol for freshness tracking
• Optionally uses GPT-5.6 for semantic judgment
• Operates fully local with no network required
• Uses MCP over stdio (JSON-RPC) transport
• Runs on Windows, macOS, and Linux

The system is described as having a "Retrieval Core", "Structural Indexer", and "Optional Intelligence" components.

Evidence: The author's own write-up describes the architecture in detail.

The description states Mnemex positions itself around:

• Distinguishing "violation vs. evolution" in code decisions
• Blocking contradictions while allowing legitimate refactors
• Using content hash anchoring instead of "vibes"
• Operating fully local with no telemetry
• Being "deterministic tagged constraints" that can block violations without models

The author claims this is a guardrail system where:

• Deterministic logic owns policy
• The model only makes semantic judgments
• It prevents "fresh anchored decision" contradictions
• It allows compatible refactors like extracting idempotency checks into helpers

Evidence: Self-reported claims in the project write-up.

The description states Mnemex targets:

• AI coding agents (specifically those using tools like Claude Code, Cursor, VS Code)
• Developers working with Python and TypeScript/TSX
• Users who want guardrails for agent code changes
• Teams wanting to prevent contradictions in automated code modifications

It's positioned as a tool for "developer experience" that works with existing editors.

Evidence: The author mentions integration with Claude Code, Cursor, VS Code, and describes developer-focused features like one-command setup.

Not evidenced. The description does not contain any information about pricing, monetization, or business model.

The description states:

• Built with SQLite, FTS5, BM25, FastMCP, Python, Node.js, OpenAI
• Uses deterministic code selection and bounded evidence from models
• Has hard token caps (800/400/800) for retrieval
• Runs fully local with no telemetry or cloud dependencies
• Supports cross-platform CI builds across Linux, macOS, Windows
• Includes a 60-second offline demo
• Installation is idempotent and byte-identical on rerun

Evidence: The author's own technical write-up.

Not evidenced. There is no mention of users, customers, revenue, adoption, or usage metrics in the description.

Not evidenced. The description does not reference any competitors or market positioning beyond its own claims.

• Unproven claims: The author states Mnemex "blocks a fresh contradiction while allowing legitimate evolution" but provides no evidence of this working in practice
• Limited scope: Only supports Python and TypeScript/TSX languages
• Hackathon project: Submitted to OpenAI hackathon, suggesting it's experimental rather than production-ready
• No commercial traction: No customers, revenue, or adoption data provided
• Model dependency: Despite being "fully local", it optionally uses GPT-5.6 which may contradict the no-cloud claim
• Unverified claims: The author explicitly states this is a "recorded-fixture replay" not a live-agent benchmark

Inference: The project appears to be an experimental proof-of-concept rather than a mature product.

1. What specific real-world scenarios have you tested Mnemex against?
2. How does it handle edge cases where content hashes don't clearly indicate contradiction vs. evolution?
3. Have you validated the "confidence 0.96" threshold in actual agent workflows?
4. What's the performance impact of content hashing on large codebases?
5. How do you plan to scale beyond Python/TSX support?
6. What are the actual use cases where this system would be valuable vs. just a theoretical improvement?