The first post in this series covered five ways AI memory fails. This one goes deeper on the quietest of them: losing track of what changed.
Why isn't searching your transcripts enough?
Because a transcript records everything that was said, not what's true now. Search finds every passage that matches your question, including the ones that have since changed, and leaves your AI to guess which is current. That guess is where answers go wrong.
Picture three meetings over a quarter. In the first, the budget is $500K. In the second, the team talks about raising it. In the third, it's settled at $750K. Ask your AI "What's the budget?" and a transcript search returns all three passages, each a strong match for the question. Which one should it believe? A careful person would check the dates. A search result doesn't come with that judgment built in.
What does a dated fact give your AI?
It says what was true and when. Edwin stores every fact with the date it was true, so when a budget moves from $500K to $750K it records an update, not an error, and the old value stays on record with its date. Your AI gets the current answer and the history behind it.
Dates do more than settle the latest value. They let a memory answer the questions people actually ask about change: what was the plan in March, when did it move, and what did it replace? A transcript holds those answers too, scattered across hours of conversation. A dated fact holds them in one line.
Decisions benefit in the same way. Edwin keeps the decisions from your meetings as top-priority memories, tied to the meeting they came from and the facts discussed with them.
How much more accurate are dated facts?
About twice as accurate in our test. Turning meetings into dated facts answered 25.6% of questions correctly, against 13.1% for searching the raw transcripts, with the same answering model in both cases (pre-registered, Oct 2026).
The test used STALE, a third-party benchmark built on facts that change over time: the exact situation this post is about. Both approaches used the same AI model to write the answers. The only difference was what that model was given to read: passages from the raw transcripts, or the dated facts Edwin had made from the same meetings.
An earlier internal test points the same way. With source dates, Edwin classified 50 of 50 updates correctly; without them, 23 of 50 (internal test, Feb 2026).
What can you do about it today?
Whether or not you use Edwin, three habits help any AI keep up with change:
- Date what you tell it. "As of October, the budget is $750K" beats "the budget is $750K."
- Say what replaced what. "We moved the launch from May to June" gives your AI the change, not just the latest value.
- Keep decisions with their meeting. A decision is easier to trust when you can see where it was made.
Edwin does this for you. It turns your meeting transcripts (Otter.ai exports today) and the notes your AI clients save over MCP into dated facts, so the date travels with every fact automatically.
For technical readers: how was it measured?
The comparison was pre-registered: the questions, the two approaches and the success measure were written down before the run. Both used the same answering model, so the gap reflects the memory, not the model. The results, with their confidence intervals, are in our October 2026 measurements.
Dates tell your AI when a fact was true. The other half of trust is knowing whether it's been checked, which is where the labels from the first post come in.