Mnexium AI

Product Introduction

1. Definition: Mnexium AI is an AI memory infrastructure layer that provides persistent, explainable context management for large language models (LLMs). It operates as a middleware between applications and LLM APIs.
2. Core Value Proposition: It eliminates manual context management by automatically storing, scoring, and recalling long-term user context—solving LLM amnesia without vector databases or custom retrieval pipelines.

Main Features

1. Chat History Persistence:
   • Logs every message within sessions via log: true flag.
   • Automatically injects prior conversation context using history: true, maintaining continuity without manual window management.
2. Agent Memory Engine:
   • Extracts and stores structured user facts/preferences (e.g., dietary restrictions) with semantic embeddings.
   • Assigns dynamic importance scores (0-100) to memories and auto-recalls them (recall: true) in relevant future sessions.
3. Agent State Tracking:
   • Manages short-term task states (e.g., pending bookings) via key-value pairs (state: {load: true, key: "current_task"}).
   • Enables resumable workflows across sessions.
4. Observability Hub:
   • Provides audit trails for memory triggers, showing why specific memories were recalled (e.g., "Used: mem_abc123 (preference, score 92.5)").
   • Logs all API calls and memory modifications.

Problems Solved

1. Pain Point: LLMs reset context between sessions, forcing users to repeat information. Mnexium enables persistent user memory with zero manual retrieval logic.
2. Target Audience:
   • SaaS developers building personalized AI chatbots
   • Product teams creating multi-step AI agents (e.g., travel planners)
   • Enterprises needing auditable, tenant-isolated memory (via project_id/subject_id)
3. Use Cases:
   • Customer support bots recalling user preferences across months
   • Sales agents resuming interrupted deal negotiations
   • Healthcare apps maintaining patient history compliance

Unique Advantages

1. Differentiation: Unlike vector DBs (e.g., Pinecone), Mnexium handles end-to-end memory ops—storage, deduplication, scoring, and injection—via a single API parameter (mnx object). Competitors require separate pipelines.
2. Key Innovation: Explainable auto-recall combines semantic search with rule-based triggers (e.g., "if importance >80, inject memory") and full auditability—unavailable in vanilla vector DB solutions.

Frequently Asked Questions (FAQ)

1. How does Mnexium integrate with existing LLM workflows?
   Add the mnx JSON object to OpenAI API calls—no SDKs needed. Mnexium proxies requests, handles memory ops, and returns enriched context.
2. Is Mnexium compatible with non-OpenAI models?
   Currently supports all OpenAI models. Claude/Anthropic integration is roadmap-planned. Custom model support requires enterprise consultation.
3. How does Mnexium ensure data isolation for multi-tenant apps?
   Uses subject_id (user-level) and project_id (org-level) scoping. Memories are strictly partitioned—zero cross-tenant data leakage.
4. What constitutes a "Memory Action" in pricing?
   Any memory write (learn: true), recall (recall: true), or state update. API calls without memory ops don’t count toward limits.
5. Can I disable automatic memory creation?
   Yes. Set learn: false to manually create memories via the /api/v1/memories endpoint for granular control.