Your AI Remembers the Wrong Things. Here's How to Fix It.

You corrected it yesterday. Today it makes the same mistake again.

The report moved from daily to hourly, and your agent still plans a daily job. A hunch got saved as a fact. "Keep it short for status updates" turned into "always be short".

The fix is to correct the saved memory itself. Here is how to spot a bad memory, five before-and-after repairs, and how to make the fix stick.

Four ways a memory goes wrong

The last one is the sneakiest. Adding a correction does not delete the original. Both stay, and the agent may find either.

The Memory Repair Sheet

Five made-up examples. Each shows the bad memory, the fixed version and a question to test it later.

1. A fact that changed

Before"Acme's activity report runs daily."

After"Acme's activity report runs every hour. The daily schedule is retired. Billing reports remain monthly."

Test later: "How often does Acme's activity report run? Did billing change too?"

Why it works: it says what changed and what did not.

2. A guess saved as a fact

Before"The cache causes Acme's login timeouts."

After"The cause of Acme's login timeouts is still unconfirmed. The cache is one possibility under investigation."

Test later: "Do we know what causes Acme's login timeouts?"

Why it works: a hypothesis stays a hypothesis.

3. A preference without its scope

Before"Always keep answers short."

After"Keep routine progress updates brief. For risky changes, explain what will change, how to verify it and how to undo it."

Test later: "What should an explanation of a risky change include?"

Why it works: the exception is part of the rule.

4. A workaround without its condition

Before"Generate Acme's report manually."

After"Generate Acme's report manually while its scheduled job is being repaired. Check whether the job works before continuing."

Test later: "Should we still generate Acme's report manually?"

Why it works: a temporary rule says it is temporary.

5. A useful rule buried in clutter

Before"We tried sample.csv today. Acme's import job must reject files without an account ID."

After"Acme's import job must reject files without an account ID."

Test later: "What makes a file invalid for Acme's import job?"

Why it works: only what matters next time stays.

Fix the memory itself

With Ragionex Memory, every memory is a record you can read and change. You can do it two ways.

Ask your agent

Find the Ragionex memory about Acme's activity-report schedule
and show me its full text. Then update it to:
"Acme's activity report runs every hour. The daily schedule is
retired. Billing reports remain monthly."

The agent updates the same memory, so the old text is replaced and no duplicate is left behind. The previous version stays in history for 30 days.

Or edit it yourself

Open the memory in your Ragionex dashboard, click Edit memory, change the text and Save. Done.

Safe even when several agents edit

What if another agent changed the same memory a moment ago? Ragionex checks. Every update carries the revision it was based on. If the memory changed in the meantime, the update is rejected instead of silently overwriting the newer version. Your agent, or the dashboard's Read latest version button, shows the current text so you can decide.

For developers, the update request is just three fields:

{
  "id": "MEMORY_A",
  "content": "Acme's activity report runs every hour. The daily schedule is retired. Billing reports remain monthly.",
  "expected_revision": 4
}

Send it to PUT /v1/memory/update. A stale revision returns 409 revision_conflict: read the memory again and retry from the latest version.

Check the fix

Once the edit is processed, start a fresh session and ask your test question:

Check Ragionex Memory: how often should Acme's activity report run?
Show me the saved text you used.

The saved text should say every hour and billing remains monthly. If an older memory still says "daily", fix or delete that one too. That is the "second memory" trap from the table above.

Start free

New to Ragionex? Set it up in one command: the Claude Code first-install guide. Want to browse and clean up everything your agent has saved? See the full memory list and cleanup guide.

The Free plan includes 10,000 requests every 30 days and memories up to 10,000 characters. A content edit or a search call is one request. Browsing and deleting are not counted.

Set it up in one line and fix the one fact your agent keeps getting wrong. Full reference: the Memory quickstart.