Let AI Teach Me · Make it stick

Let AI Pick Up the Last Chat in a New Thread

The context window is this session’s desk. Close the chat and it is empty. The layer you want to keep has to be written outside.

“Traveling with Dad, 67. At most three places a day. Back at the hotel by 4 p.m. to rest.”

That was the constraint you sent in round 3. Saturday afternoon at the table you were plotting a five-day family trip, then hotels and transit. It kept picking up the thread, so you treated the line as settled: I already told it. It should remember.

By round 18 you asked for the final itinerary. Day four had five places, last item an 8:30 p.m. show. You did not check line by line. You pasted it to the family chat. Dad wrote back: “Wasn’t that day supposed to be back by four?”

You judged that way because, among people, a confirmed request becomes shared ground. AI’s replies also stayed coherent, as if it had followed from the start. What it depends on is text it can still read and still attend to. That earlier line may have been pushed away, or dropped. Said once is not the same as enforced every time after.

AI can pull out the agreed constraints, the decisions, and why options were dropped. That copy still has to live outside the chat, and get pasted back at the next start. Where is the gap? Watch the same trip plan get continued.

Why it remembered a moment ago, then dropped it

Think of the context window as the papers AI has spread on the desk right now. A destination just mentioned is still on the desk, so it can keep talking. The desk has an edge. The longer you talk, the harder it is for earlier details to stay in the answer. Open a new thread and the new desk will not bring the old constraints with it.

A transcript still in the UI only means you can find those words later. That is not the same as “AI will treat this line as a constraint next time.” What must be kept for the long run cannot live only in one chat’s history.

Have AI pick out what should be kept

When you are about to close an important chat, have AI sort it into the four buckets below. It finds. You check. Then store the result in notes or a project file.

  1. Goal: what this should finish, for example a five-day itinerary that works for a family of three.
  2. Hard constraints: what every later answer must obey, for example at most three places a day, hotel by 4 p.m.
  3. Decisions already made: what you picked and why, for example staying by a metro stop to cut transfers.
  4. Drops, and why: what you tried and why it is off the table, so it does not come back next time.

Why it was dropped is the line people skip. Write only “no evening shows” and next time AI may recommend a different show. Write “Dad needs rest in the afternoon, no evening events” and it knows the boundary.

Try it first · Three ways to continue, and how they differ
A half-hour trip-planning chat
Half an hour later, how you plan to continue
All three choices start from the same trip plan.
Wait until the window fills, then tap and see.
1 · Talk through a half-hour trip plan firstWatch the window fill
2 · Then end or stretch this chatSee whether the early constraints are still there
3 · Then look at AI’s next lineSee which conditions it keeps using
What AI can use for this answer
How full the window is now0%
After play, what you just talked about appears here.
The four-line agreement you saved outsideAt most three places a day / hotel by 4 p.m. / no more than one metro transfer / no evening shows, Dad needs rest
What AI says next (schematic)
After play, AI’s next reply appears here.
Pick a way to continue, then play the full run.
What happens under each of the three waysA new chat starts with a new window. In a long chat, early details may leave the current pile. An agreement saved outside can be pasted back in.
The step that actually decides whether it can pick upBefore the next answer, did it read again the agreement that has to be kept.
Teaching demo: window fill and replies are simplified.
Paste this straight to AI

The block below can sit at the end of an important chat. It only sorts. It does not decide where you store it.

Copy this whole block
  1. Set the scope firstReview this chat and make a record for next time. Write only what already appeared. Do not fill in guesses.
  2. Then fix four columnsOutput four columns: goal; hard constraints for every later answer; decisions already made and why; options dropped and why.
  3. Have it keep the numbers and your wordingKeep concrete numbers for people, time, budget, and counts. Quote my wording on key constraints when you can. Mark anything unsure as “to confirm.”
  4. Have it write the next opening pasteFinally write a paragraph I can paste at the next start, so the hard constraints can be checked fast and none are dropped.

After it writes, check the numbers and the drop-reasons, then store them outside. If AI copies “before 4 p.m.” as “early evening,” the next itinerary will still be wrong.

The second line’s why it was dropped is the hinge. A new chat then knows why a choice is out, and the same boundary can block a similar suggestion. Names without reasons quickly become a list you cannot reuse.

The next chat starts from these four lines

We are still planning a five-day trip for a family of three. Dad is 67. At most three places a day. Back at the hotel by 4 p.m. No evening shows, Dad needs rest at night. Restate these constraints first, then give day four.

That opening puts the easiest-to-lose conditions from the old chat back into the current window. A reliable start for long-term memory is text you can save, check, and read in again.

Where that text lives, and how to keep it in layers, is the next page: external memory in three layers. Remember one thing here: a transcript can be reread. A long-term agreement needs its own save.

If you want the next chat to pick up, write the agreement outside first.