Agent Memory Privacy — Research Notes

A living repository of research reports, paper notes, and ideas on privacy-preserving memory for AI agents: machine unlearning, contextual integrity, differential privacy, and memory-architecture defenses.

Repository structure

.
├── reports/      # Long-form research reports & surveys (dated, bilingual where available)
├── papers/       # Per-paper reading notes (one file per paper)
├── ideas/        # Research ideas, open problems, project proposals
├── templates/    # Templates for paper notes and idea write-ups
└── RESOURCES.md  # Benchmarks, key groups, venues, living link lists

Reports

Date	Title	Languages
2026-06	Privacy-Preserving Memory for AI Agents: State of the Art and Open Problems	EN · 中文

Conventions

• Report filenames: YYYY-MM-topic.<lang>.md (e.g., 2026-06-frontier-survey.en.md).
• Paper notes: papers/YYYY-firstauthor-shorttitle.md, using templates/paper-note.md.
• Ideas: ideas/NNN-short-title.md, using templates/idea.md. Number sequentially so ideas are easy to reference in discussion.
• Reports are treated as living documents; substantive revisions get a changelog entry at the bottom of the file.

Scope

Four pillars (see the frontier survey for the full taxonomy):

1. Machine unlearning / knowledge editing — removing information from model parameters; verification of forgetting.
2. Contextual integrity — CI-aware information flow control for what agents disclose, store, and retrieve.
3. Differential privacy — for training, fine-tuning, in-context learning, RAG, and memory retrieval.
4. Memory-architecture defenses — admission control, TTL/retention, selective forgetting, deletion verification, multi-agent channel controls.

Disclaimer

Notes here summarize and paraphrase published work for research purposes; see each file's references for original sources. Claims from arXiv preprints are the authors' and not independently verified.