Vault Agent Memory | Agent Memory Governance

Start by giving this to your Agent.

If you already work with agents, you should not have to learn forty Vault commands first. The first path is agent-led: answer four questions, then read a short daily report.

Install Vault Agent Memory for this project. Use vault-for-llm[mcp]==0.7.31.
Use the agent-assisted governed-auto memory mode.

Do not show advanced CLI flags first. Ask me only four questions:
1. Which language should Vault use: Traditional Chinese, Simplified Chinese, or English?
2. Should this be an independent vault or a shared vault for multiple agents?
3. Should Vault connect to Obsidian, Supabase, both, or neither?
4. What time should the daily memory report run?

After setup, run a smoke check and tell me:
- where the vault lives
- how I read the daily memory report
- where the local GUI or next action is

Daily rule:
safe, low-risk, sourced memories can be kept automatically;
uncertain, sensitive, conflicting, or strategic memories should go into the daily report for my review.

One repo. Three Agents. One governed memory.

This is the demo that separates Vault from a plain RAG database: agents can share lessons without silently polluting long-term memory.

1. Propose

Claude Code fixes a testing bug.

It proposes a reusable lesson with the command output and source file that proved the fix.

2. Review

Vault does not trust it blindly.

The candidate passes privacy, duplicate, metadata, and quality checks before it can become active memory.

3. Reuse

Codex finds it next time.

Search returns a reviewed memory, then bounded read gives the exact source range for citation.

Without Vault

Every agent starts cold.
Fixes stay trapped in chat logs.
Bad memories can silently enter context.

With Vault

Agents share reviewed project memory.
Each answer can return to a source.
Outdated memory can be deprecated or rolled back.

Open the visual demo Read the technical runbook

The Daily Memory Report keeps humans in control.

Vault can become more automatic over time, but the review surface stays small. Safe memory can enter the vault; uncertain memory waits for a person.

Auto-kept

Testing requires APP_ENV=test before pytest.

Reason: verified by command output. Source: CI log and test run.

Needs review

Vault should focus on memory governance.

Reason: strategic product direction. Action: approve, edit, or defer.

Rejected

Today's market snapshot.

Reason: short-lived data, not durable project memory.

Obsidian conflict inbox

Your notes stay readable. Agent memory stays governed.

When Obsidian and Vault both changed a note, the resolver gives clear choices instead of overwriting silently.

Accept Obsidian Accept Vault Keep both

Adapters can change. The governed memory layer stays yours.

Vault starts local-first, then connects to the systems you choose. CLI, MCP, Gateway, Supabase, Obsidian, and future remote servers are adapters around the same memory lifecycle.

Codex

Claude Code

Hermes

OpenClaw

n8n

Coze

Vault Gateway / MCP / CLI

Local SQLite Vault + governed memory lifecycle

Obsidian

Supabase

Remote Server

Search QA

Daily Report

Audit Trail

Developer proof without making the first page heavy.

The landing page tells the story. The repo keeps the proof: local SQLite, MCP profiles, candidate-first writes, Obsidian import/export, Supabase and Gateway adapters, Search QA, and Apache-2.0.

Local-first

Core use stays on Markdown and SQLite. Remote sharing is optional.

MCP-ready

Start agents with small tool profiles instead of exposing every maintenance tool.

Measurable

Search QA checks whether agents find the right source, not whether they sound confident.

Reviewable

Candidates, daily reports, conflict inboxes, and audit trails keep memory accountable.

pip install "vault-for-llm[mcp]"
vault quickstart
vault demo agent-governance --json