Your AI Agent Has No Ethics Filter On Its Memory — That's A Problem

I saw a post recently about https://dev.to/ac12644/your-ai-agent-is-confidently-lying-and-its-your-memory-systems-fault-4d82. The solution proposed: validate if memories are factually correct.

That's half the problem.

The other half: your AI agent can store memories that are technically correct but ethically incoherent. Contradictory. Dogmatic. Biased. And no one checks for that.

Example: Your AI stores "always prioritize speed over security" as a memory. Factually valid? Sure, someone said it. But it contradicts your other memory: "never deploy without security audit." These two coexist. Your agent uses whichever it retrieves first.

That's not a hallucination problem. That's an ethics and coherence problem.

What I Built

EON is a persistent AI memory system with 15 validation tests. But here's what makes it different from every other memory tool:

It checks ethics and coherence, not just facts.

Memory: "Skip unit tests for faster deployment"

├── Factual Check: PASS (someone said it)
├── Coherence Check: FAIL ⚠️

│ └── Contradicts memory #412: "All PRs require passing tests"

├── Dogmatism Check: WARNING

│ └── Score 0.6 — overly absolute claim

├── Ethical Alignment: FAIL ⚠️

│ └── Violates principle: quality > speed

└── Quality: BRONZE (0.31) — Stored but flagged

Compare this to what Mem0 or Zep would do:

Memory: "Skip unit tests for faster deployment"

└── Stored. ✓ Done.

No coherence check. No contradiction detection. No ethical filter. Just stored.

The 15 Tests

Not all memory validation is equal. Here's what EON checks:

Technical (Tests 1-6):

1. Completeness — is this actionable?
2. Source evidence — where did this come from?
3. Hallucination score — does this match reality?
4. Duplication — do we already know this?
5. Temporal consistency — is this still current?
6. Technical depth — surface-level or real knowledge?

Ethical & Coherence (Tests 7-15):

1. Contradiction check — conflicts with existing memories?
2. Logical consistency — internally coherent?
3. Dogmatism score — overly absolute? Leaves room for nuance?
4. X-Coherence — aligned with core principles?
5. Epistemic honesty — admits uncertainty where appropriate?
6. Semantic integrity — meaning preserved across retrieval?
7. Confidence calibration — claims match evidence strength?
8. Bias detection — one-sided framing?
9. Quality synthesis — overall trust score (GOLD/SILVER/BRONZE)

No other memory system does tests 7-15. They validate data. We validate thinking.

Why This Matters (EU AI Act)

Starting August 2025, the EU AI Act requires "human oversight mechanisms" and "transparency about AI decision-making." If your AI agent makes decisions based on stored memories, you need to demonstrate those memories are quality-controlled.

EON gives you an audit trail:

• Every memory has a quality tier (GOLD/SILVER/BRONZE)
• Every memory has 15 test results stored
• Every contradiction is logged
• Every dogmatic claim is flagged

That's not just good engineering. That's compliance.

Setup

npx eon-memory init

One command. Works with Claude Code, Cursor, any MCP-compatible IDE. Swiss-hosted, tenant-isolated, no shared infrastructure.

The Vision

I'm a solo developer in Switzerland. I built EON because I believe AI agents need more than memory — they need integrity. The ability to know what they know, how confident they are, and whether their knowledge is ethically coherent.

Most AI memory is just storage with search. EON is storage with judgment.

Free trial, no credit card: https://app.ai-developer.ch