AI Memory Vault: Turning an Obsidian Vault Into Real AI Memory

Most agent memory tools ask an operator to trust a new piece of infrastructure: a vector database, an encrypted store, a background server they now have to maintain and monitor. AI Memory Vault takes a different starting point. It treats an Obsidian vault, the note-taking app many operators already keep open every day, as AI memory: a folder of plain markdown files, wikilinks, and a few conventions that a Claude Code agent reads and writes directly. There is no database, and in the default setup, no server at all.

What AI Memory Vault actually builds

The repository is a single build script, ai-memory-vault.md, written as a skill file rather than documentation. Loaded into Claude and run interactively, it interviews the operator about their work and projects, then constructs the vault structure around the answers: a root index that orients the agent at the start of every session, folder indexes mapping each area of the operator's world, daily notes that generate themselves, a living profile that updates as the agent learns more, and a system the maker calls "Jobs," one master note per recurring task that bundles the skill for doing that task plus links to exactly the notes it needs. The pitch is specific: read one note, have the whole job, instead of re-explaining context at the start of every session.

Obsidian vault as AI memory: two files, not one

The design splits memory into two layers, and understanding the split is most of understanding the system. CLAUDE.md is a short boot config that lives outside the vault, in whatever working directory Claude Code launches from. It carries only the agent's identity, the vault's location, and rules that cannot lapse, kept short on purpose so it survives context compaction. VAULT-INDEX.md lives inside the vault itself and holds the operator's profile plus the map of the vault's folders and rules; it gets read at boot but can be compressed away in a long conversation, which is exactly why the durable rules sit in the other file instead. A third template, MEMORY.md, drops into Claude Code's own project memory folder and redirects that native memory back into the vault, so an operator doesn't end up maintaining two memory layers that quietly drift apart from each other.

Setup: filesystem access, not an MCP server

The recommended path needs no MCP server, no tunnel, and nothing new to run in the background. Claude Code, launched from a working folder next to (not inside) the vault, reads and writes the vault's files directly over the filesystem, after a one-time file-access approval. Claude Desktop users can wire the same access through the standard @modelcontextprotocol/server-filesystem MCP server instead, and a remote MCP connector is documented for claude.ai web and mobile access, though that path means exposing a local vault through a public tunnel, more setup and more surface area than most operators will want.

The honest requirement underneath all three paths: this only works for someone already running Obsidian, or willing to install it specifically for this. It is not a drop-in for an operator with no interest in a note-taking app. Every template ships with [FILL IN: ...] markers meant to be interviewed out by the AI rather than typed by hand, which speeds setup, but the system still assumes a real first session spent building the vault before it does anything useful.

Where it fits next to deja-vu, HUMAN.md, and Compartment

AIOS Guide has now covered four distinct approaches to the same underlying problem, an agent that forgets everything between sessions, and each solves it a genuinely different way. A HUMAN.md-style file is a single hand-written context document, authored once, read fresh each session. deja-vu indexes the session logs an agent already writes to disk and searches them lexically, no embeddings, built for developers re-finding a past coding fix. Compartment is a live, encrypted, embeddings-based semantic memory store the agent writes to and queries automatically.

AI Memory Vault is none of these. It is not a single file, not a log index, not a database at all: it is a structured, multi-file, human-browsable vault built on an app the operator likely already has open, organized with the same folder-index-and-wikilink conventions Obsidian users already apply to any second brain, just aimed specifically at what an AI agent needs to find and how fast it needs to find it. The tradeoff is close to the mirror image of Compartment's: nothing here is encrypted or access-controlled beyond ordinary filesystem permissions, and the memory is exactly as durable, and exactly as messy, as the operator's own note-taking discipline turns out to be.

The license, and other honest limits

The scan's "open-source" label needs a correction. The repository ships under Creative Commons Attribution-NonCommercial-ShareAlike 4.0, not a standard software license like MIT or Apache 2.0. In plain terms: free to use, share, and adapt with attribution, but not to resell, repackage, or build a commercial product directly on top of, and any derivative has to carry the same license forward. Building a business and running it on this system is explicitly fine, per the README's own plain-English summary; reselling the system itself, or a lightly modified version of it, is not. That is a materially different deal than the Apache 2.0 terms covered in AIOS Guide's Compartment piece, worth knowing before anyone plans to package this into a paid offering.

The other limit is more practical than legal. The whole system depends on the operator, or the interviewing AI, doing a genuinely good job describing their work, projects, and recurring tasks in that first setup conversation; a thin or rushed interview produces a thin vault, since there is no separate ingestion step pulling in outside knowledge the way an embeddings store might. There is also no built-in encryption, tamper detection, or access control. Everything sits in plain markdown on disk, protected only by whatever file permissions and backup discipline the operator already has in place.

The verified numbers

As of today, the AI Memory Vault GitHub repository shows 484 stars and 106 forks, with the maker still actively pushing updates. It is a legitimate pick for an operator who already treats Obsidian as a second brain and wants the AI reading from the same place they do, rather than a separate, invisible store they can never open and read for themselves.