VaultMem: Operator-Blind Encrypted Memory for Personal AI Agents

Read full text Summarise with VIRAT
Authors: Avinash Gosavi
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We present VaultMem, an embeddable Python library implementing Layer 1 of the Memory-as-Asset (MAS) framework with a cryptographically enforceable operator-blind storage guarantee. Every memory atom is encrypted client-side with AES-256-GCM before leaving session RAM; the platform developer stores only ciphertext and cannot read user memories regardless of their privacy policy. We introduce the .vmem vault format -- a self-describing, tamper-evident, language-agnostic binary standard for portable encrypted AI memory -- and specify a formal session-scoped key architecture with bounded MEK exposure and O(new atoms) append-only checkpointing. We survey existing AI memory systems against a four-dimension privacy taxonomy and show no prior system achieves all four properties simultaneously. We characterize the embedding privacy problem as a two-phase challenge and provide a five-level mitigation hierarchy. Three-tier retrieval (AFFINITY->COMPOSITE->ATOM) achieves MRR=1.00 on pattern queries versus MRR=0.50 for flat cosine, with Recall@10=0.84 and zero regression on specific queries. The prototype is ~2,100 lines of Python with 39 unit tests, available at https://github.com/aag1091-alt/vaultmem-sdk.

Fields of Study & Subjects

Cryptography and Data Security Privacy-Preserving Technologies in Data Security and Verification in Computing

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
21 March 2026
Language
EN
Document type
Preprint