Product
Vaultic
A persistent memory server for any MCP-compatible AI agent. Transcripts, searchable memories, a knowledge graph and a relationship ledger, all in one local file. No cloud.
Linux only. Windows is coming soon. macOS is not tested. Please check this before you buy.
- No subscription, pay once
- All 0.x updates included
- 14-day refund, see the refund policy
The price includes any applicable tax. Checkout is run by Lemon Squeezy, the merchant of record; we never see your payment details.
A few of the 85
session_bootStart a session and get the agent’s profile and recent context back.log_turnSave every conversation turn word for word, append-only.rememberStore a distilled memory, with near-duplicates merged.recallFind memories by meaning, with filters.retrieve_contextPull the right context from memories, the graph and the session in one call.correct_ledger_entryCorrect a belief without erasing it. The old claim is kept, marked refuted.record_commitmentTrack a promise until it is kept or broken.get_continuity_capsuleA snapshot of active state, so a fresh session can pick up where the last one stopped.
…and 77 more, covering the knowledge graph, the relationship ledger, consent, retention, export and the audit log.
Different room.
Same history.
Switch clients, models or tools and the agent’s memory stays with them. It lives in one local file that any MCP client can reach.
See how it works- Any clientDifferent room.
- Any modelDifferent room.
- Any sessionDifferent room.
- One memory. Many clients.
- Conversations become structured knowledge.
- Identity, preferences, context and history persist.
- You stay you. It all stays on your device.
Not just storage.
Structured memory.
Conversations, corrections and commitments become a connected knowledge graph your agent can search.
See what you get- Several memory types
- Knowledge graph of entities and relations
- Relationship ledger
- Corrections refute, never overwrite
- Evidence and provenance
- Hybrid keyword and semantic retrieval
Built for evidence,
not guesswork.
A plain pipeline from raw conversation to trusted context.
Capture
Every turn saved verbatim, append-only.
Provenance
Each belief links back to its evidence.
Consolidation
Near-duplicates merge. Corrections refute the old claim.
Retrieval
Hybrid keyword and semantic search.
Runtime context
The right memory, handed over when asked.
Original evidence stays intact. Interpretations can evolve.
What you get
Local-first
Everything lives in a single SQLite file on your own device. No cloud component, no telemetry and no account.
Encrypted at rest
Memories, transcripts, reflections and observations are sealed with AES-256-GCM before they touch disk, so a copied database or backup holds ciphertext.
Personas
Give each agent its own persona and memory. Scope memories by user when a persona is shared, and let an agent keep a working or personal relationship with its user.
Nothing silently overwritten
Corrections refute the old claim and add a new one. Forgetting archives a memory instead of erasing it. Commitments and repairs are tracked the same way.
You stay in control
Per-user consent categories, an audit log of every read and write, and diagnostics that are off unless you turn them on.
Works with your client
Standard MCP over stdio, with HTTP available, and 100 tools. Your agent is shown a short list of the 31 it needs every day and reaches the rest through a built-in menu, so they do not use up its working space. Point any MCP-compatible agent at it.
What it stores
Four kinds of data, each with its own purpose.
| Kind | What it is |
|---|---|
| Transcripts | The raw, verbatim conversation record. Append-only. |
| Memories | Distilled facts, preferences, summaries and events, found again with hybrid search and de-duplicated automatically. |
| Knowledge graph | Named entities, typed relations and free-text observations. |
| Relationship ledger | What the agent believes about the user, itself and the relationship, each claim linked to its evidence. |
Security and privacy
- Content is encrypted at rest by default with AES-256-GCM.
- The key store sits in its own directory with restrictive file permissions. There is deliberately no passphrase, so the server can open its own memory on every start.
- No telemetry, and no licence server. The software never phones home to check your licence or look for updates.
- The embedding model is included in the package, so it works fully offline from the first run.
- Diagnostics are opt-in, and contain only aggregate counts and version strings, never your content.
- Using an external embeddings API is an explicit choice you make. It is not the default.
The full detail is in the privacy policy.
What you need
- Python 3.12 or newer, whose SQLite is 3.51.3 or newer. The Python that ships with Ubuntu 24.04 (3.12) is too old. The install steps use
uvwith Python 3.13, which brings a current SQLite. The server refuses to start on an older one, because older SQLite has a bug that can corrupt the database. - An internet connection while installing, to download the dependencies. The embedding model is included, so after installing it runs fully offline.
- Developed and tested on Linux. Windows support is coming soon. macOS has not been tested.
Honest limitations
We would rather you know these before you buy.
- Back up your encryption keys. The key folder lives apart from the database, and an encrypted backup cannot be read without it. Your data and keys never leave your machine, so we cannot recover them for you.
- Keyword search over encrypted content matches whole words only. There is no stemming, prefix or phrase search. Recall combines it with semantic (meaning-based) search, and the database holds only keyed fingerprints of the words, never the words themselves.
- The relationship safety check is a pattern list, not a semantic classifier. A sufficiently different phrasing can slip past it. It is a safeguard, not a guarantee.
- De-duplication, decay and contradiction flagging are heuristics. They are not verified understanding.
- It is built for one local user. It is not a multi-tenant system, and user scoping is bookkeeping, not authorisation.
- File-permission hardening is tested on Linux only.
We tested our own recall claims against a public benchmark — see the results.
Getting started
After purchase you download the package from your order, and you can re-download it at any time. Install it with pip using the step-by-step instructions included, then register it once, globally: add the full path of the vaultic-mcp command inside the environment you installed it into to your client's user-level (global) MCP settings, not to a single project or session. If it is only added to one session, your other sessions start without it and cannot reach your memory. Registered globally, every session in every folder shares the same memory:
{
"mcpServers": {
"vaultic": {
"command": "vaultic-mcp"
}
}
}Your download also includes an Agent Manual, written to your AI agent. It covers installing and connecting, the rules for every turn, all 100 tools, triggers, hooks and skills, and comes with a rulebook skill and reference hooks for Claude Code.
Updates arrive by re-downloading from your order. The software never checks for them itself.
Questions first? Get in touch.