Zycaskevin Vault For LLM Vault For LLM

Vault-for-LLM Project Memory

Use this skill when OpenClaw needs durable project knowledge from Vault-for-LLM.

Vault is not an always-on chat memory dump. Treat it as governed project memory: confirmed decisions, source-of-truth notes, pitfalls, SOPs, repo docs, and reviewed handoff knowledge.

Agent Rules

1. Search before answering project-memory questions.
2. Do not treat search preview text as final evidence.
3. If a result has id, node_uid, or line_start/line_end, use bounded read before citing it.
4. Cite the source title and line range returned by Vault when making claims.
5. Use candidate-first memory:
   • For new observations, use vault_memory_propose.
   • Do not promote candidates unless a human operator explicitly asks.
   • Do not use direct vault_add for autonomous memories.
6. Keep secrets out of memory. If content may include tokens, API keys, private credentials, or customer data, ask before storing.

Commands

vault-openclaw init
vault-openclaw search "deployment checklist" --limit 5
vault-openclaw read-range 12 --line-start 20 --line-end 55
vault-openclaw mcp-config

Database Scope

OpenClaw shares memory with other agents only when they use the same Vault project directory. If OpenClaw should join a shared project memory, configure the plugin projectDir to the same path used by Hermes, Codex, Claude Code, or n8n. Use a separate projectDir for isolated experiments.

Recommended Loop

vault_search -> vault_read_range -> answer with source

For new memory:

vault_memory_propose -> human review -> vault_memory_promote only after approval

Source: zycaskevin/Vault-for-LLM — distributed by TomeVault.