What we've learned building Edwin, a memory layer for the AI tools you already use: how to tell a checked fact from a confident guess, how to read a memory benchmark, and how to keep your memory yours. The full measurement record is on our research page.
Showing all 5 posts
AI memory goes wrong in five predictable ways, from trusting confident guesses to losing track of what changed. Here is each failure, why it matters, and the design choice Edwin makes to prevent it.
A model's confidence tells you how thoroughly it analyzed a claim, not whether the claim is true. Edwin keeps those two questions apart and labels every memory with what has actually been checked.
Published AI memory scores often move when someone else re-runs them. Here is what to look for in a benchmark claim, the questions to ask any vendor, and what the public record shows, dated and linked.
A second AI is worth having when it is genuinely different from the first and its verdict reaches the answer as a clear label. What we learned building Edwin's cross-provider checks.
What your AI knows about your work is worth keeping. Edwin stores it as a file on your own Mac, backs it up with encryption every night, and works with any MCP-compatible AI client.