Your AI agent is brilliant for exactly one session. Then the context window closes, and it forgets everything: what you told it, what it learned, what it promised to do next.

I got tired of re-explaining myself every time I opened a new chat. So I built Vaultic — an MCP server that gives any MCP-compatible agent a memory that actually persists.

What it stores

Vaultic keeps four kinds of data, each with a different job:

Transcripts — the raw conversation record, append-only. Your agent can search its own history by phrase, date, or role.

Memories — distilled facts, preferences, events. Not just dumped text — Vaultic runs hybrid search (semantic + encrypted keyword) and de-duplicates automatically. When your agent remembers something it already knows, it updates instead of creating a duplicate.

Knowledge graph — named entities with typed relations and observations. Your agent builds a structured understanding of the people, projects, and systems it works with.

Relationship ledger — what the agent believes about you, itself, and the working relationship, with every claim linked to its evidence. Corrections don't silently overwrite — they refute the old claim and record the new one.

108 tools, but your agent only sees 31

One problem with big tool sets is context window bloat. Vaultic solves this by advertising 31 core tools by default — the ones an agent needs every day: recall, remember, reflect, commit, repair, session lifecycle.

The other 77 tools (graph analysis, retention policies, backup/restore, conflict detection, temporal queries, artifact storage) are reachable through a built-in menu tool. Your agent calls list_tool_catalog() to see what's available and call_tool() to use any of them. Full capability without filling the context window.

The five packs

Pack Tools Covers
core 31 Transcripts, memory CRUD, recall, knowledge graph, reflections, commitments, repairs, sentiment, working state, session lifecycle
memory 18 Consolidation, extraction rules, agent profiles, user profiles, tags, intentions, learning proposals
graph 12 Graph reads, temporal queries, evidence provenance, episode tracking, conflict detection, contamination checks
artifacts 10 File/image storage with versioning, experience short-term memory
admin 37 Backup/restore, retention policies, consent management, boundaries, search index, diagnostics

Local-first, encrypted, no cloud

Everything lives in a single SQLite file on your machine. No cloud component, no telemetry, no account, no licence server. The software never phones home.

Content is encrypted at rest with AES-256-GCM before it touches disk. The embedding model ships with the package, so after a one-time install (needs internet for pip dependencies), it runs fully offline.

The keyword search index uses keyed fingerprints of words — the actual words never appear in the index. So even the search layer doesn't store your plaintext.

How it compares

The memory space has a few established players. Here's where Vaultic sits:

Vaultic Mem0 Letta (MemGPT) agentmemory
Hosting Local only Cloud Cloud + self-host Local
Price €135 one-off $0–249/mo $0–200/mo Free (OSS)
Encryption AES-256-GCM at rest Cloud-managed Cloud-managed None
Knowledge graph Built in Pro tier only Not built in Not built in
Relationship tracking Built in (ledger) No No No
Tool count 108 ~10 ~15 ~5
Telemetry None Yes Yes None

The trade-off is real: Vaultic doesn't have a managed cloud offering or a free tier. It's a one-time purchase, you run it yourself, and it's Linux-only right now (Windows is coming).

Setting it up

After purchase, install with pip (the included instructions walk you through it with uv and Python 3.13), then register it globally in your MCP client settings:

{ "mcpServers": { "vaultic": { "command": "vaultic-mcp" } } }

Register it globally — not per-project — so every session in every folder shares the same memory. That's the whole point: your agent remembers across contexts.

Your download includes an Agent Manual written to your AI agent. It covers all 108 tools, the rules for every turn, triggers, hooks and skills, plus a ready-made skill file and reference hooks for Claude Code.

Honest limitations

I'd rather you know these before you buy:

  • Back up your encryption keys. They live apart from the database, and we can't recover them.
  • Keyword search matches whole words only. No stemming or prefix search. Semantic search covers the gap for most queries.
  • The relationship safety check is a pattern list, not a semantic classifier.
  • De-duplication and decay are heuristics. Good heuristics, but not verified understanding.
  • Linux only right now. Windows is coming. macOS hasn't been tested.
  • Single-user. User scoping is bookkeeping, not authorisation.

We tested our recall claims against a public benchmark — the results are on our site.

Why I built it

I needed my agent to remember things. Not in a "retrieval-augmented generation" way where you bolt a vector database onto an LLM. In an actual structured way — where it tracks what it's committed to, admits when it was wrong, maintains a model of who I am and what I care about, and does all of this in a format it can reason over.

Everything else in the space was either cloud-hosted (my data leaves my machine), subscription-based (I pay forever for what's essentially a database), or too simple (just key-value memories with no structure).

So I built what I needed, and now it's available.

Links

Questions? Get in touch.