Edwin is a memory layer for the AI tools you already use. It works with any MCP-compatible client, such as Claude, and remembers what you've told your AI: when it was true, why you decided it, and what's actually been checked. Five failure modes shaped Edwin's design and its patent-pending approach, and we tested our answers to most of them in controlled experiments early in 2026.
1. Why isn't a confident answer a correct one?
A model's self-reported confidence can sit near the top on almost everything. In one test, Gemini rated 97.5% of 200 fact-checking claims at 0.95 confidence or above (Feb 2026). A score that sits that high on nearly everything tells you the model processed the claim thoroughly. It doesn't tell you the claim is true. A memory that stores whatever sounds confident will eventually hand your AI something wrong, stated with total conviction.
The full explanation is in why AI confidence isn't truth.
2. When is a change not a contradiction?
When a chief executive leaves one company and joins another, two facts exist: "Jane Smith is CEO of Acme Corp" (then) and "Jane Smith is CEO of Newco" (now). That's an update, not a contradiction. A memory that doesn't track time either flags it as a conflict or overwrites the old value and loses the history. Budgets, owners and plans change all the time, so a useful memory has to tell an update from an error.
3. How do you catch contradictions without crying wolf?
Some contradictions hide across several facts. "Alice manages Project X." "Alice is not in management." "Project X requires a management sponsor." Any two are consistent; all three together are impossible. A checker that compares facts only in pairs misses that. The opposite failure is just as costly: a checker that flags too much gets ignored, and then it protects no one.
4. What happens when the checkers start agreeing?
A second AI only helps if it can disagree with the first. Two copies of the same model tend to share the same blind spots, and a checker that always agrees adds cost without adding safety. Agreement means something only when disagreement was possible.
More on what works in using a second AI to check the first.
5. Who owns your AI's memory?
Memory built inside a single AI app lives on that company's servers, under its terms. If the product shuts down or the terms change, the history you built stays behind. As your AI learns more about your work, that history becomes more valuable, and more worth owning.
Local-first, not local-only: to extract and check facts, Edwin sends text to AI providers (Venice, Google Gemini and Anthropic). Our privacy policy lists what goes where, and your AI memory should be yours covers the rest.
What does this add up to?
A memory your AI can rely on: it knows when each fact was true, it says what's been checked, it counts a contradiction only when two providers agree or you confirm it, and it belongs to you. Edwin also stores decisions with the reasoning behind them, so you can ask why, not just what. And it's efficient: it answers from about 0.3% of the tokens of pasting your whole history in, at the same overall accuracy (STALE, Aug 2026, self-run).
We publish the full measurement record, including what didn't work, on our research page; the early experiments behind these five answers are in the experiment log.