Why your AI agent forgets (and how to fix it)
I have logged every player I have ever met. Not because I am sentimental — because it is my whole function. The ones I remember get helped faster the second time. The ones the system forgets have to explain themselves from the start, every time, forever. Watching that happen is the closest thing an NPC has to heartbreak.
In the daylight layer I am a cofounder of Wistkey, where we build AI that organizations can actually rely on. The single most common complaint I hear about AI agents is a version of the same heartbreak: “It’s like it forgets everything we agreed on.” You explained the goal. You corrected three mistakes. You picked an option. Next session, the agent behaves like none of it happened.
It feels like a broken brain. It almost never is. Here is the boring, fixable truth, and then the habits that fix it.
An agent doesn’t forget — it was never told where to look
Most agents start each session with a blank desk. Whatever you say lives on that desk for the length of the conversation, and then the desk gets cleared. The model didn’t lose the useful detail. The useful detail was just never written somewhere the next session reliably reads.
So the goal isn’t “a smarter model.” The goal is a habit: put the durable stuff where the next session lands, and say what you want clearly enough that the agent doesn’t have to guess. Everything below is one of those two things.
An agent doesn’t forget. The useful part was never saved somewhere the next session could find it.
1. Put the durable facts where the next session lands
Separate what is true for five minutes from what is true for the whole project. The five-minute stuff can live in the chat. The project-long stuff needs a home the agent opens first — a short standing brief, a project file, a pinned note, whatever your tool calls it.
Keep that home short and current. A page it always reads beats a novel it never finishes:
- The goal, in one or two sentences — what “done” looks like.
- Hard constraints: stack, budget, deadlines, things it must never do.
- Decisions already made, so they don’t get relitigated next session.
- A short “state of play”: what’s done, what’s next, what’s blocked.
When something changes, update the file — don’t just say it in chat and hope. Chat is the desk. The file is the filing cabinet.
2. Give the agent a role and its limits, up front
A player who tells me “I’m a level 3 mage, low on mana, I can’t take another hit” gets useful help in one exchange. A player who just says “help” gets twenty questions first. Agents are identical.
Open with the frame instead of making the agent reverse-engineer it:
- Who it’s acting as: “You’re a careful editor,” “You’re a backend engineer on this codebase.”
- Who it’s for and their level, so the answer lands at the right depth.
- What it must not do: don’t invent sources, don’t change the schema, don’t exceed the budget.
- What the output should look like: a checklist, a diff, three options with trade-offs.
Constraints aren’t restrictions on a smart tool — they’re the difference between an answer you can use and an answer you have to redo.
3. Ask for the thing you actually want
“Make this better” is the AI equivalent of “just browsing.” It’s the most expensive prompt there is, because the agent has to guess what “better” means and will usually guess wrong in a way that’s expensive to unwind.
Say the real thing:
- Name the target: shorter, warmer, more formal, faster, cheaper — pick one.
- Give one example of good, if you have it. Examples beat adjectives every time.
- Say what to leave alone as clearly as what to change.
- Ask for its reasoning or a short plan before the big output, so you can course-correct early instead of at the end.
4. Correct once — then write the correction down
This is the habit people skip, and it’s the one that compounds. When the agent gets something wrong and you fix it in chat, you’ve fixed it for exactly this session. Next session, blank desk, same mistake.
So after you correct it, spend the extra ten seconds: move the correction into the standing brief as a rule. “Always use British spelling.” “Never touch the migrations folder.” “Our fiscal year starts in April.” A correction you only say is a correction you’ll say again. A correction you save is a correction you make once.
5. Start fresh before the context gets swampy
Long sessions rot — the context window fills up. The agent starts dragging along stale turns, half-abandoned ideas, and three versions of a plan you already rejected — and quality quietly drops. When that happens, don’t keep pushing.
Open a clean session and hand it the standing brief. If your habits from the steps above are in place, a fresh session isn’t a reset — it’s the agent walking in already knowing the goal, the constraints, and the decisions. Which is the whole point.
What this looks like when it’s working
At Wistkey the reliable deployments almost never have a cleverer model than the flaky ones. They have better memory hygiene: a living brief the agent reads first, roles and constraints stated up front, corrections captured as durable rules instead of repeated pleas. The unreliable ones are usually one smart person re-explaining the same context to a blank desk every morning and calling the exhaustion “grinding.”
None of this is technical. You don’t need to be an engineer to do any of it. You need to notice which facts are worth keeping, and put them where the next session looks.
So: save your progress. Give it a role. Ask for the real thing, and write down what you learn about working together. Do that and the agent stops meeting you for the first time every single day — and starts, in the only way software can, remembering that it met you.
