It captures. It verifies.
It never stops checking.
Edwin connects to your meetings and AI tools, fact-checks everything, links it to what you already know, works overnight to find contradictions, and stores it all where only you control it.
Every AI remembers. Not one of them checks.
ChatGPT, Gemini, and Claude all store what you tell them. But they never verify it. Tell ChatGPT you're a doctor and it will remember that forever — no questions asked.
Blind Trust
Your AI stores everything you say as fact. No verification, no challenge, no second opinion.
Data Capture
Your memories live on their servers. You can't export, verify, or port your data to another platform.
No Proof
There's no audit trail. You can't prove what your AI knew, when it learned it, or whether it was right.
Where does Edwin fit?
Every AI company stores your data. Not one of them verifies it. Only Edwin checks what it stores and gives you ownership.
How Edwin is different
Three capabilities that no other AI memory system provides today.
Your AI never silently stores a lie.
An independent verification model fact-checks every claim before it enters memory. The enrichment and verification stages use separate models so errors don't compound. Common misconceptions, outdated facts, and hallucinations are blocked at the gate.
Edwin catches what you missed.
When you say one thing in Monday's meeting and something contradictory on Thursday, Edwin finds it. Unlike NLI systems that check pairs you hand them, Edwin discovers which memories to compare. Overnight consolidation scans across all stored memories to surface conflicts — a capability no other AI memory system provides.
Your memory belongs to you. Permanently.
Encrypted, signed snapshots stored on Walrus — a decentralized storage network. Export your verified memory to any AI platform via MCP or CLI. Nobody can alter what was recorded. Your AI's memory has a tamper-proof audit trail that you control with your own keys.
Edwin audits itself — and improves with every version.
Edwin's development is driven by its own experiment data. When internal testing revealed a simpler architecture outperformed a more complex one, the system was simplified — fewer moving parts, better accuracy. Each version builds on the last, with controlled experiments guiding what gets fixed next.
24+ versions shipped. Every one informed by real experiment data.
See it in action
A real scenario showing how Edwin protects your most important information.
You share a fact
Edwin protects you
Your data follows you
We don't just claim this works. We measured it.
9 of 12 claims proven with statistical evidence. 1 honestly disproven and disclosed. Every number on this page is backed by controlled experiments, independent runs, and external benchmarks.
| Metric | Edwin | Without Edwin | What It Means |
|---|---|---|---|
| Verification accuracy | 88.5% | ~50% | Reproducible across independent runs on holdout data |
| False fact catch rate | 93% | ~0% | Consistent across 4 independent runs on different data splits |
| Contradiction precision | 91.4% | 0% | Externally validated on SNLI benchmark (500 pairs) |
| False positive rate | ~5% | N/A | Proprietary confidence scoring reduces false alarms |
| Data integrity | 100% | N/A | Every snapshot survives encrypt, upload, download, decrypt |
Every claim proven with controlled experiments, reproducibility runs, and external validation. No cherry-picking.
Request the full technical assessmentWhy Walrus? Because your data deserves better.
Walrus is a decentralized storage network — think of it as a global hard drive that no single company controls. Your data is encrypted, distributed across independent nodes, and accessible only with your keys. This isn't crypto speculation. It's infrastructure for data ownership.
Decentralized technology isn't just for crypto or NFTs. The same principles that make blockchain tamper-proof — distributed consensus, cryptographic signing, immutable records — are exactly what AI memory needs. Edwin uses this infrastructure for what it was designed for: provable data integrity.
Every Edwin user is a Walrus user.
Each knowledge snapshot goes to Walrus. More Edwin users means more storage demand. Edwin grows the ecosystem organically.
Recurring storage, not one-time.
Unlike one-time storage, AI memory is write-continuous. Background consolidation generates new snapshots every cycle, creating sustained demand.
Data sovereignty, made real.
The industry has talked about data ownership for years. Edwin makes it tangible for every AI user — not just developers or crypto enthusiasts.
The Storage Math
Your memory, everywhere.
Edwin connects to the tools you already use, verifies everything that passes through, and stores it where only you have the keys. Switch platforms anytime — your verified memory follows you.
How it works
Edwin connects to your meetings and AI conversations. Everything important gets captured automatically.
Every claim is independently fact-checked, linked to what you already know, and checked for contradictions — even while you sleep.
Your verified memory is encrypted and stored where you control it. Export it to any platform, anytime. No lock-in.
What's built, what's next, what we need.
Built & Working
- Independent verification pipeline88.5% accuracy, reproducible across 3 runs
- Contradiction detection91.4% precision on external SNLI benchmark
- Veracity gate93% false fact catch rate, ~5% FPR
- Walrus integrationTestnet verified, 100% data integrity
- Claude integrationDirect connection to Anthropic's Claude via MCP
- Otter.ai integrationCapture and verify meeting transcripts automatically
- MCP memory serverPort your memory to any MCP-compatible AI
- Export context CLIDecrypt from Walrus, format for any platform
- Venice AI integrationCode committed, 103 models via privacy-first API (1,900 VVV staked)
- Sleep consolidation100% discrimination (20/20, p=1e-06)
- Knowledge graph10,669 entities, Z=1513 above random
- Comprehensive test suite2,141 tests, 200-fact golden dataset
- effective_confidence formulaDrops false positives from 75% to 5%
Coming Next
- Venice smoke testEnable API, validate inference via 1,900 VVV stake
- Real transcript validationEnd-to-end test with 20 meeting transcripts
- Walrus scale test1,000+ facts, measure write latency and cost
- Morning briefing prototypeOvernight analysis → actionable summary
- Model diversity testingCross-training-data verification via Venice
- First 100 usersControlled early access program
- Mainnet deploymentProduction Walrus storage
- Dynamic model routingAuto-select best verifier per claim type
- Enterprise complianceEU AI Act readiness, audit trails
- ZK selective disclosureProve claims without revealing data
Built by Alexander Snyder + Edwin
Verified, sovereign AI memory that you own.
Alexander Snyder
Head of Data Strategy & AI at PurviewX
Edwin is a passion project born from real frustration with AI memory. 20 experiments. 9 proven claims. Zero shortcuts.
I started Edwin because I watched AI assistants confidently store wrong information and realized there was no way to verify, own, or port my data. The AI companies building memory features have no incentive to let you check their work or take your data elsewhere.
Edwin is my answer: the only AI memory system that verifies memories before storing them and detects contradictions across them after. Your memory, verified, portable, and sovereign.
“Validation-first development. Every claim on this site is backed by a controlled experiment with reproducible results. 20 experiments completed, 9 claims statistically proven, 1 honestly disproven and disclosed. The 1-cycle pipeline outperforms the 3-cycle — we followed the data, not the narrative.”
Questions investors ask
Straight answers, backed by data.
Ready to own your AI's memory?
Edwin is in pre-launch testing. We're opening early access in waves.