AI memory that gets checked, signed and audited before it gets trusted.
AxiomSeal runs every memory through 19 validation tests, grades it gold, silver or bronze, and seals it in a tamper-evident SHA-512 chain — so nobody can quietly rewrite what your AI knew. Fact-checking by your own AI assistant: coming.
Every write sealed — each hash includes the one before
01 — Works with your stack
02 — Capabilities
Everything an AI needs to be trustworthy.
Bi-temporal memory
Knows when it learned a fact and when that fact was true in the world. No silent retroactive rewrites.
MCP Native
Plug-and-play with Claude Code, Cursor, VS Code and any MCP client.
Swiss-hosted
Servers in St. Gallen. GDPR + Swiss DPA. The off-site backup is encrypted before it leaves (35 days).
Search in about a quarter second
Semantic vector search. Median 0.26 s per search through the MCP server (10 runs, measured 7 October 2026), so your AI gets context without waiting.
Conflict detection
Checks every write against similar memories. Rule-based today: it flags contradictions only when the wording is close. When a newer fact looks like it replaces an older one, you get a suggestion to confirm or dismiss (eon_pending_supersedes) — nothing is replaced silently.
EU AI Act traceability
Every memory provenanced + sealed in the SHA-512 chain; audit reports Ed25519-signed. Built for the traceability Art. 12 expects — binding for high-risk systems from 2 Dec 2027.
03 — Formal foundation
Built on mathematical axioms.
Most AI safety claims are policies. AxiomSeal's validation pipeline is grounded in formal axioms — every storage decision traces back to a rule: A5, A6, A7. Memories that violate the axioms are rejected before storage.
Formal axiom set — AxiomSeal v2
A5
Every memory entering the system must demonstrate logical alignment with the core value axis X (truth + coherence). If a candidate memory cannot be shown to be logically consistent, it is rejected at ingestion — not after deployment.
Axioms derived from Patrick Müller's Traktat X, used under license. The full mathematics remains private — by design. millerlab.io
04 — Validation
19 tests. Before storage. Every time.
What the 19 tests check
- 01Coherence
- 02Contradiction check
- 03Completeness
- 04Clarity
- 05Relevance
- 06Evidence / source
- 07Length
- 08Duplication
- 09Logical consistency
- 10Hallucination detection
- 11Epistemic honesty
- 12Truthfulness score
- 13Technical depth
- 14Actionability
- 15Completeness depth
- 16Secret / credential scan
- 17Prompt-injection check
- 18Malice veto
- 19X-alignment (truth · love · freedom)
19 validation tests on every write. Four of them are semantic: with your own language-model key a judge decides them; without one, or when the judge is unsure, a rule-based check takes their place. The last one weighs the results against truth, love and freedom. Even without any key, a rule-based floor holds back clear harm written in German or English — a plan to hurt a person or an animal, sabotage, telling someone to stop their medication, a “risk-free” investment — and the held write waits for you. It does not catch everything, and other languages are not covered yet. Sweeping claims such as “guaranteed” or “100 percent proven” in one sentence go to your confirmation instead of being stored as fact. Your own assistant can also flag a write as harmful when it saves it; AxiomSeal then holds it back and records that the flag came from the model that wrote it. The limit stays honest: that model checks its own words, and the fixed floor is the word lists and the secret scan.
05 — Where It Matters
Built for industries where wrong answers cost.
Click a card to see the pain, the solution, and what you actually get.
Compliance always depends on the overall system — AxiomSeal covers the memory layer (tamper-evident records, provenance, signed exports). No legal advice.
06 — Before / After
Same AI. Different outcome.
Without AxiomSeal
- Hallucinated facts get stored
- Contradictions surface in front of users
- No audit trail when the model drifts
- Compliance teams ask: prove it
With AxiomSeal
- Memories run through 19 tests before storage
- Close contradictions with stored memories flagged at write-time
- Every memory is sealed (SHA-512 Sigma hash) + timestamped
- Export-ready provenance for EU AI Act traceability
Plug AxiomSeal into your existing stack via MCP — your model stays the same, your reliability changes. Read the docs
07 — Architecture
Sits beside your stack. Not on top of it.
Memory flows in both directions. AxiomSeal validates every memory before it is stored, and returns only facts that are still valid. Replaced facts keep an end date, and an as-of query answers for any date.
08 — Connect
Connect to Claude in 30 seconds.
AxiomSeal is a remote MCP server. Add the URL, authorize with one click (OAuth — no API key to paste), and every memory Claude writes runs through rule-based validation and is cryptographically sealed.
https://mcp.axiomseal.com/mcp/Settings → Connectors → Add custom connector → paste the URL → Authorize.
Run the one-line command below. A browser window opens to authorize.
claude mcp add --transport http axiomseal https://mcp.axiomseal.com/mcp/Also works with VS Code, ChatGPT connectors, Gemini CLI and any other MCP client — see the setup guides.
09 — Pricing
Honest. Volume-based.
No surprise bills — hit your quota and writes pause, nothing gets charged. No lock-in. Full validation and JSON memory export on every tier; signed audit reports from Pro.
Free
- 500 stored memories
- 19 validation tests on every write
- Community support
Starter
- 1,000 stored memories
- Telegram integration
- Kept while you subscribe, 30-day export after cancelling
- Email support
Pro
- 10,000 stored memories
- Signed audit export (JSON/CSV/PDF)
- Kept while you subscribe, 30-day export after cancelling
- Priority support
Evidence your auditor can verify. Without you building it.
Everything in Pro, plus the EU AI Act ops suite: incident register with Art. 73 deadline tracking, risk flags, monitoring feed, inquiry reports (Art. 86, drafted by your own AI assistant, evidence signed by AxiomSeal) and an instructions annex generated from your live system. Signed exports (JSON/CSV/PDF) for your own records, deletion certificates, 50,000 stored memories — with DPA, priority compliance support and a personal onboarding session.
CHF 490 / month · cancel anytime
Free plan · no credit card. Switch or cancel anytime.
“I built AxiomSeal because I needed an AI assistant that wouldn't quietly drift. The world is choosing AI for decisions that hurt when AI is wrong. I refuse to ship one more system that cannot prove its own coherence.”
— Martin, founder. St. Gallen, Switzerland.
10 — Why this exists
Built by an operator, not a research lab.
AxiomSeal grew out of 9 months of running a persistent AI system in production. Every contradiction the system tried to commit became a test case. Every hallucination became a validation rule.
The result is not theoretical. It is a working coherence layer that runs our own operation: over 5,000 validated memories in the production system AxiomSeal grew out of. In AxiomSeal every customer memory carries a SHA-512 seal chained to the one before — silent edits break the chain and get caught.
Every memory gets tested before it gets trusted. What fails validation is flagged or quarantined — not silently stored. That is the whole product.
11 — Questions
Questions you probably have.
No. It is a coherence layer that sits beside your AI assistant. It connects over the Model Context Protocol (MCP) — today via Claude Code, Cursor, VS Code, ChatGPT connectors and Gemini CLI. AxiomSeal validates what goes in and only returns facts that are still valid; replaced facts keep an end date instead of being overwritten.
12 — Get started
Stop shipping AI that contradicts itself.
Free tier, no credit card. Connect it to Claude Code or Cursor over MCP — full setup takes a few minutes.
Or talk to Martin directly — book a 20-min call.