"Saved isn't searchable: how to check your agent actually has your team's context"
A saved meeting is not the same as one your agent can find. A five-minute checklist to test that your team's context is connected, processed and retrievable.
A meeting being saved does not mean your agent can find it. To check that your agent really has your team's context, test four things in order: the source is connected, it has been processed, it is searchable, and the agent retrieves it with a source.
This is the most common way agent memory quietly fails. Everything looks fine. The call is recorded, the email is imported, the tool says "success". Then you ask a question and the agent answers confidently without the fact you know is in there, or with a fact it cannot back up.
Why this fails silently
Between "saved" and "answered" there are several steps, and any of them can stall without an error message:
- The source never connected, or only part of it did.
- The content was received but not yet processed into something searchable.
- It was processed, but the search looks in the wrong place or under the wrong permissions.
- The agent found it but answered without telling you where it came from.
An empty answer looks the same whether nothing exists or the search looked in the wrong place. That is why you need to test, not assume. This applies to any memory setup, whether you built it yourself or bought it.
The checklist
Run this with one meeting or thread you know well, ideally something from this week.
1. Is the source connected?
Confirm the thing you expect to be captured is actually feeding in.
- Was the meeting bot invited to that call, or is the call one of those you chose to record?
- Is the right inbox connected, and is it the account where the thread lives?
- Is the calendar connected for the people on that call?
If the source was never connected, nothing downstream can fix it. This is also the easiest problem to miss, because nothing fails; there is just no data.
2. Is it processed?
Raw transcripts and raw emails are not the same as usable knowledge. Processing is what turns them into something an agent can search, such as facts about people, companies and decisions.
Check that your recent item has been through that step. Allow some time after a long call or a large import. If your tool shows a status for each source, use it. If it does not, move on to step 3, because the test there tells you the same thing.
3. Is it searchable?
Search for something specific that you know is in that call or thread. Use a distinctive detail, such as a number, a name, or a phrase someone used, not a general topic.
- If you find it: good. It is indexed.
- If you do not: do not assume it is missing. Try the same search with the person who attended, since results depend on whose permissions apply. A no-result answer can mean "does not exist" or "you are not allowed to see it" or "the search looked in the wrong place".
4. Does the agent retrieve it, with a source?
Now ask your agent, not the search box. Ask a question that depends on the fact, and add: "Tell me which meeting or message this comes from, and quote the relevant line."
A good answer includes the fact, the date, the source, and a quote. A weak answer includes the fact alone. A bad answer includes a fact that is not in the source at all.
Then run a control. Ask the same question in a second agent connected to the same source. Both should return the same fact. If one does and one does not, the problem is in a connection, not in the content.
Quick diagnosis
| What you see | Likely cause |
|---|---|
| Agent finds nothing, and nothing was ever recorded | Step 1: source not connected |
| Item exists but cannot be found by detail | Step 2 or 3: not processed or not indexed yet |
| Teammate finds it, you do not | Permissions, or a different account |
| One agent finds it, another does not | The second agent's connection |
| Fact appears with no source | Memory without provenance; do not rely on it |
Make it a habit
Run the four steps once when you set up, again after you connect a new agent, and any time an answer surprises you. It takes about five minutes and saves the much worse discovery that your agent has been working from partial context for weeks.
How this looks in Nysa
Nysa turns a team's meetings, email and calendar into facts with sources: quotes, dates and where each came from, so step 4 has something to check. Agents read the same brain over MCP, and each sees only what its teammate can see, so a search from a different person can return different results, as described in step 3.
To run the checklist: connect your agent with Connect Claude or Connect ChatGPT, pick a call from this week, and ask for the fact and its source. For the underlying concepts, see what is shared memory for AI agents and how to give every agent on your team the same company context.
FAQ
How long should I wait before testing a new meeting?
There is no fixed answer that holds for every tool. Test a short while after the call ends. If the item is not searchable, try again later, and if it is still missing after a reasonable wait, work back through steps 1 and 2.
What does a good source look like?
A date, the meeting or message it came from, and a quote or link you can open. If an agent cannot give you at least that, you cannot tell a real decision from a guess.
Why does a teammate get different results than I do?
Because access is scoped per person. With Nysa, each agent sees only what its teammate can see, and raw email stays private to its owner. Different results can be correct behavior, not a bug.