People expect a bot to remember like a person. It doesn't. A bot's memory is closer to an office's memory: it's whatever got written down and filed.
Three kinds of memory
| Kind | What it is | Lasts | Example |
|---|---|---|---|
| Working memory | The context window, what's on the desk now | This session | The email you just pasted |
| Saved memory | Notes the bot keeps about you and its job | Until changed | "The owner works in Central time" |
| Files and skills | Documents the bot reads and writes | Until changed, with versions | A plan file, a skill file, a record |
The model itself doesn't learn from your chat. When the session ends, working memory is gone. What carries forward is saved memory and files.
Compaction: the quiet filter
Long sessions fill the desk. When that happens, the system summarizes older parts and drops detail. That's compaction.
Most of what it drops is noise. But nothing guarantees it won't drop something important. Say you told the bot early on, "never email the customer directly." After enough rounds, that line can get summarized away. Now the bot is working without the one rule that mattered.
That's why What is a bot? sets the bar at "come back tomorrow." A bot has to survive compaction and the end of the chat.
How a good bot handles memory
- It writes important things down. Rules, decisions and preferences go in a file or saved memory, not just the chat.
- It asks before it saves a rule. "Should this be the new default?" turns a one-off correction into standing memory, with your yes.
- It re-reads before it acts. It opens the plan file instead of trusting a summary.
- It keeps core memory short. Short core memory, details in skills. A long memory is harder to keep accurate.
That last one comes from a review of 90 published marketplace bot templates, where nearly all kept core memory short and pushed detail into skills (our notes, good-practice.md).
The ISO parallel
Organizations solved this long ago. People leave. People forget. So you keep organizational knowledge in controlled documents and records, not in someone's head.
A bot's chat is like someone's head. Its files are the organization's memory. See Knowledge that survives the night.
A real example
Our AN Creation office bots have a "save" skill. When a record is finished, it's saved to a set folder, with a dated copy kept in History. Nobody has to hope the bot remembers what was decided last Tuesday. It's in the file.
What this means for you
- If it matters tomorrow, ask: "Where is that written down?"
- Review saved memory now and then. Old preferences go stale.
- Don't paste the same rule into chat every day. Put it in a skill once.
- When a bot "forgets," don't scold the chat. Fix the file.
Memory you can open and read is memory you can manage.