My support bot stopped forgetting everything when I changed AI memory write triggers

A support bot initially performed poorly in real-world scenarios due to a failure to remember relevant information. The bot would repeatedly ask customers to reboot routers, confirm firmware versions, and engage in other repetitive troubleshooting tasks, despite having access to a large context window and a GPT-5.4 model.

The root cause was that the bot was storing every customer utterance as part of its long-term memory, which included irrelevant and redundant information such as "yes," "still broken," and "I already tried that." This approach polluted the retrieval process, introduced repeated mistakes, and unnecessarily bloated prompts with low-value context.

The author of the story discovered that the issue stemmed from the timing of when memory was written. Initially, memory was written on every customer interaction, but this resulted in a cluttered and noisy memory system. To address this, the author implemented a set of five memory write triggers that focused on storing relevant information only:

1. Confirmed facts – Store facts only when they are explicitly confirmed by the customer. For example, store the customer's ISP and device firmware when these details are confirmed, but do not store speculative information like "probably has an old router."

2. Preference changes – Remember customer preferences, such as their preferred communication channel, restart policies, and language style. These preferences should be stored as they directly impact the support interaction.

3. State-changing troubleshooting steps – Record only those troubleshooting steps that result in a change of system state, such as a factory reset, DNS change, or firmware rollback. Avoid storing suggestions that are not actual state changes.

4. Decisions – Store significant decisions made during the support interaction, such as escalating the issue to a higher tier or scheduling a technician visit. These decisions influence the next steps in the workflow.

5. Outcomes – Document the final outcome of the support interaction, whether the issue was resolved, escalated, or closed. Outcomes provide closure and prevent the bot from reopening resolved cases.

By implementing this focused memory write strategy, the support bot's retrieval process became smaller, cleaner, and more useful. The bot no longer "forgot" prior interactions, as it only stored relevant, actionable information in its long-term memory. This approach proved more effective than attempting to store and process the entire chat transcript, ultimately leading to a more efficient and accurate support experience.