Agent Memory Vault

Overview

Use this skill to help a user install or operate the public Agent Memory Vault system. The vault can be shared by Claude Code and Codex: Markdown remains the single fact source, while each host keeps a thin rule and hook adapter. Treat the GitHub repository as the product source, and treat this skill as the Agent-facing runbook for setup, maintenance, and troubleshooting.

Repository: https://github.com/mcncarl/agent-memory-vault

Safety Rules

• Never copy a user's real memory vault into a public repository.
• Never commit .env, SQLite databases, Zvec/vector stores, model caches, logs, API keys, cookies, tokens, passwords, private chat logs, customer data, or personal absolute paths.
• When migrating an existing memory system, migrate structure, scripts, and sanitized patterns only.
• Before publishing or pushing, run leak checks and inspect the diff.
• If a task involves deleting files, ask for explicit user permission first and use Trash, not permanent deletion.
• Never point Claude Code auto-memory at the formal vault. Either disable it or treat it as non-authoritative scratch memory.
• Shared cookies and tokens belong in a private config outside the vault and Git, with owner-only permissions; Agents should inject selected values into commands instead of printing the whole file.

Runtime v2 Boundary

Before any normal command, run memoryctl --actor <actor> version --json. A managed Runtime is usable only when the response has ready=true, runtime_api_version=2, writer_protocol_version=2, a valid state-schema attestation, and a current Runtime/config/Hook integrity hash. During install or migration, normal commands must fail closed; do not bypass the transition marker with a direct Python entrypoint.