Rami's Universe

Ember AI memory research proposal

3 min read

Research Proposal: Privacy-First Hybrid Memory Architectures for AI Superconnectors

Introduction

Large-scale LLM applications are constrained by a fixed context window. This makes it difficult to recall long-term user data across evolving conversations. For Ember, an AI “superconnector” for University of Calgary students, this limitation directly impacts its core function: matching peers for mentorship, study, and friendship. Without an efficient memory store, Ember cannot reliably surface the right connections at scale.

The research question is: Which memory architecture—hybrid relational+vector, graph-based, or pure-vector—yields the highest end-to-end peer-matching accuracy for Ember while meeting cost, scalability, and privacy constraints? This question matters because solving it enables trustworthy, scalable AI systems that handle sensitive personal data responsibly.

Background

Several recent systems attempt to extend LLM memory beyond a single context window:

  • Zep’s Temporal Knowledge Graph (Graphiti) introduces temporally-aware entity linking for evolving user context [Zep 2025].

  • Mem0 focuses on dynamic fact extraction and retrieval, offering high performance at low latency [Mem0 2025].

  • MemoryOS applies an operating-system-like hierarchy of short-, mid-, and long-term storage [MemoryOS 2025].

  • HEMA uses hippocampus-inspired summaries and episodic vectors for ultra-long dialogues [HEMA 2025].

These approaches demonstrate progress but leave open questions. Pure-vector systems suffer from noisy retrieval at scale. Graph-based methods can be complex to implement and costly to maintain. No study has yet compared these approaches directly in the context of structured peer-matching where both semantic similarity and precise filters (e.g., courses, availability, interests) must be combined. This project addresses that gap.

Objectives / Hypotheses

  1. Architectural Benchmarking: Compare hybrid relational+vector, graph-based, and pure-vector architectures in terms of matching accuracy, cost, and scalability.

  2. Dynamic Memory Evaluation: Assess how quickly and reliably each system incorporates user updates and prunes stale facts.

  3. Privacy Enforcement: Test whether each architecture can enforce a strong boundary between private and public data without degrading retrieval quality.

Hypothesis: A hybrid relational+vector approach (Postgres + pgvector + structured filters) will provide the best trade-off of accuracy, scalability, and privacy, outperforming both graph-only and pure-vector alternatives.

Methodology

  1. Literature & Code Review

    • Study Zep’s Graphiti [Zep 2025], Mem0 [Mem0 2025], MemoryOS [2025], and HEMA [2025].

    • Document architectural strengths and weaknesses.

  2. Prototype Implementations

    • Hybrid Postgres: relational schema + pgvector + JSONB filters.

    • Graph-Based: deploy Zep OSS, model Ember profiles as entities and episodes.

    • Pure-Vector: PostgreSQL + Pinecone or Qdrant with metadata.

  3. Benchmarking Experiments

    • Datasets: synthetic Ember user profiles (30K–100K).

    • Metrics: precision@K for matching, latency per query, throughput, token usage per query, 12-month cost projections.

  4. Privacy Evaluation

    • Encryption-at-rest and in-transit.

    • GDPR “right to be forgotten” flow.

    • Separation of private vs. public fields using column-level encryption and role-based access.

  5. Evaluation of Success

    • Accuracy: >10% improvement in precision@K over pure-vector baseline.

    • Cost: ≤20% increase in cost compared to cheapest alternative.

    • Privacy: zero leakage of sensitive attributes in peer-facing APIs.

Outcomes

  • Comparative Performance Report: Clear benchmarks of hybrid, graph, and vector memory architectures on Ember-like workloads.

  • Design Guidelines: Actionable recommendations for AI developers building privacy-aware memory systems.

  • Prototype Code: Open-sourced implementations demonstrating hybrid triggers and retrieval pipelines in Postgres.

Impact: This work informs both the AI memory research community and practical builders of large-scale, socially embedded AI agents. For Ember, the results directly guide the development of a trustworthy, cost-efficient, and scalable student superconnector.

References

  • Zep: A Temporal Knowledge Graph Architecture for Agent Memory. arXiv:2501.13956 (2025).

  • Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory. arXiv:2504.19413 (2025).

  • MemoryOS: Memory OS of AI Agent. arXiv:2506.06326 (2025).

  • HEMA: A Hippocampus-Inspired Extended Memory Architecture for Long-Context AI Conversations. arXiv:2504.16754 (2025).

  • Zep Blog. “The Memory Foundation For Your AI Stack.” 2025.

Graph View