#セキュリティ

3 projects tagged with this

INDEX Memory Vault (IMV)

Title: Claude keeps forgetting, so I built a vault to store memories on my computer (with measurement results) A common problem everyone faces when using Claude Desktop — it forgets everything once a conversation ends. Project context, company information, rules set last week. We have to explain it all over again every time. That's why I built it: INDEX Memory Vault (IMV) — a local MCP memory vault that attaches to Claude. Three key features: 1. **Your data stays on your computer only.** No cloud, no accounts, no server transfers. It's just a single SQLite file, so you can back it up, delete it, or move it entirely. 2. **Memories have a verification status.** Stored memories can be set to verified/pending/blocked status, ensuring the AI only trusts human-approved information, rather than scraping just any memory. 3. **We measured the effects and made all data public.** 160-session A/B benchmark results — with the same model and questions, accuracy was 23.3% without the vault, and 96.7% with the vault connected. Hallucinations were 0 in both cases (if information isn't in the vault, it states it doesn't know instead of fabricating). Since these are self-measurements, you might be skeptical, so we've released all raw data, scoring rules, and answer keys. You can re-score them yourself. → [Benchmark Results Report](homepage_post_URL) Installation: If you have Claude Desktop, simply double-click the .mcpb file. → [GitHub Release v0.2.3](release_URL) It's open-source (AGPL-3.0). Please try it out and report any issues or strange behaviors on GitHub Issues — screenshots of installation roadblocks are especially welcome. That's the most valuable feedback for me. We are a two-person development company based in Incheon, building AI control structures. We built this believing that memory is not about storage, but about retrieval and moderation.

1132

jong kimJul 7, 2026

#AI Policy#Security#Social Issues

Idea

The Importance of Agent Governance

When an agent's memory is leaked or corrupted, its actions become compromised. Agents perform critical tasks like sending emails, processing payments, and managing files, all based on their stored memories. Memory storage is evolving from a mere database into the foundational ledger for an agent's actions. Consequently, I estimate an over 80% probability that 'memory poisoning' will emerge as the primary attack vector in the era of AI agents. This underscores why describing privacy as 'life' is not an exaggeration. Without a strong focus on privacy, agents operating in this data-rich environment are bound to cause incidents. A fundamental lack of governance exists. Developers must collaborate, research, and share insights to establish the essential governance protocols.

172

jong kimJul 6, 2026

#AI Policy#Security#Social Issues

Idea

AI Control Governance Idea

Develop specific ideas for control and governance solutions aimed at the safe and ethical operation of AI systems. What specific problems do you want to focus on?

172

jong kimJul 5, 2026

#AI Policy#Security#Social Issues

PPOMNAE