SuperLocalMemory: Privacy-Preserving Multi-Agent Memory with Bayesian Trust Defense Against Memory Poisoning

2026-03-04 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Software Development & Engineering · Depth: Expert, extended

Summary

SuperLocalMemory is an open-source, local-first memory system for multi-agent AI designed to defend against OWASP ASI06 memory poisoning. It achieves this through architectural isolation, Bayesian trust scoring, and adaptive learning-to-rank, all without cloud dependencies or LLM inference calls. The system integrates SQLite-backed storage with FTS5 full-text search, Leiden-based knowledge graph clustering, and an event-driven coordination layer with per-agent provenance. Its adaptive re-ranking framework learns user preferences via three-layer behavioral analysis, improving NDCG@5 by 104%. Evaluation on an Apple M4 Pro demonstrates 10.6ms median search latency for 100 memories, zero concurrency errors with 10 simultaneous agents, and 72% trust degradation for "sleeper attacks." Behavioral data is isolated in a separate database, supporting GDPR Article 17 erasure requests. SuperLocalMemory is MIT-licensed and integrates with over 17 development tools via the Model Context Protocol.

Key takeaway

For AI Architects or CTOs evaluating memory solutions for multi-agent systems, SuperLocalMemory offers a compelling local-first, privacy-preserving alternative to cloud-dependent systems. Its built-in Bayesian trust defense against memory poisoning (ASI06) and adaptive, zero-LLM personalization directly address critical security and relevance concerns. Consider integrating this MIT-licensed system to enhance data control, reduce attack surfaces, and improve agent performance in sensitive developer workflows, especially where GDPR compliance is paramount.

Key insights

Local-first AI memory with Bayesian trust and adaptive re-ranking defends against poisoning and personalizes retrieval without cloud or LLMs.

Principles

• Local-first architecture enhances security and privacy.
• Bayesian trust models detect gradual adversarial behavior.
• Adaptive re-ranking improves search relevance without LLMs.

Method

SuperLocalMemory employs a four-layer memory stack (storage, hierarchy, graph, patterns), an event coordination layer, and an adaptive learning layer. It uses Beta-Binomial Bayesian inference for trust scoring and a three-phase re-ranker (baseline, rule-based, ML) for personalization.

In practice

• Use SQLite with WAL for concurrent read access.
• Isolate behavioral data for GDPR compliance.
• Implement provenance tracking for forensic analysis.

Topics

• Multi-Agent Memory Systems
• Memory Poisoning Defense
• Bayesian Trust Scoring
• Adaptive Learning-to-Rank
• Local-First Architecture

Code references

• varun369/SuperLocalMemoryV2
• topoteretes/cognee

Best for: AI Architect, CTO, VP of Engineering/Data, AI Engineer, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents — AI agents struggle to discriminate subtle, context-dependent relations (complementary, nuanced, contradictory) among long-term memories.
• From Data Models to Mind Models: Designing AI Memory at Scale — Agentic memory enables AI systems to learn, adapt, and retain knowledge, crucial for complex, stateful operations.
• Designing Memory for AI Agents: Inside Linkedin’s Cognitive Memory Agent — LinkedIn's CMA provides a stateful memory layer for AI agents, enabling continuity, personalization, and adaptation across interactions.
• MRMMIA: Membership Inference Attacks on Memory in Chat Agents — Chat agent memory is vulnerable to membership inference attacks, which MRMMIA effectively demonstrates across various access levels.
• Beyond Similarity: Trustworthy Memory Search for Personal AI Agents — Semantic similarity in AI agent memory retrieval creates critical trustworthiness failures; task-conditioned memory admission is essential.
• CLASP: Defending Hybrid Large Language Models Against Hidden State Poisoning Attacks — CLASP defends Mamba-based LLMs from hidden state poisoning by classifying malicious tokens via block output embeddings.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.