Practice: Memory and reflection
For education only. This practice is about agent architecture; it is not investment, financial, or trading advice, and nothing here is a recommendation to buy or sell anything.
Self-check
Section titled “Self-check”Six short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. Why does the system record a decision and reflect on it at two separate moments instead of one?
Show answer
Because you cannot judge a decision the moment you make it: the outcome is not known yet. Recording is instant and cheap; judging requires hindsight. So the system writes the decision down right away (marked not-yet-judged) and waits until enough time has passed to see how it actually turned out before reflecting.
2. What does the word “pending” mean on a logbook entry, and what replaces it?
Show answer
Pending means the decision was made but its outcome is not yet known, so it has not been judged. It is the placeholder for hindsight. Later, once the outcome is visible, the entry is resolved: the reflection is written and the pending marker is replaced.
3. When does the reflection actually happen?
Show answer
On the next run for the same company. At the start of that run, before the pipeline goes to work, the system settles any earlier pending entries for that company. By then real time has passed, so the outcome can be measured. Entries for other companies wait until those companies are run again.
4. How is the reflection grounded in something real, rather than the system’s own opinion?
Show answer
It is fed the actual outcome: what happened over the following days (five by default), compared against a baseline of how a broad market slice moved over the same days (the framework uses the S&P 500). Measuring against a baseline is fairer than a raw up-or-down, because it asks whether the analysis added anything over what would have happened anyway.
5. Which model writes the reflection, and why does that fit the track’s pattern?
Show answer
The cheaper, faster model. Summarizing a known outcome into a few sentences is not the hard judgment (that already happened at the gate), so it does not need the most capable model. Same idea as the rest of the track: model tier follows the difficulty of the thinking.
6. Of the twelve agents, which one receives the accumulated memory, and why does that placement matter?
Show answer
Only the final gate, the portfolio manager, the agent that makes the last call. Putting the hard-won history exactly where the decision is committed is the same principle as spending the capable model there: concentrate what matters at the point of decision, not scattered across everyone.
Try it yourself: add memory to your own workflow
Section titled “Try it yourself: add memory to your own workflow”No tooling, no cost; this is design judgment. Pick an AI workflow that makes a repeated call you could later check (for example: an assistant that drafts a reply-or-escalate decision for each support ticket, or one that flags which leads to follow up). Then design its memory:
1. RECORD. When the system decides, what do you write down, and how do you mark that the outcome is not yet known?2. WAIT and RESOLVE. How much later can you actually see the outcome? What real result will you compare against (and is there a baseline that makes the test fairer)?3. REFLECT and FEED BACK. What short lesson do you store, and which step in the next run gets to read it?Show answer (worked example: reply-or-escalate on support tickets)
- Record: when the assistant decides “reply” or “escalate,” log the ticket, the date, and the choice, marked not-yet-judged.
- Wait and resolve: a few days later the real outcome is visible (was the ticket reopened? did the customer come back unhappy?). The baseline that makes it fair: how often tickets of that type get reopened anyway, so a reopen is judged against the normal rate, not in a vacuum.
- Reflect and feed back: store a one-line lesson (“escalating billing disputes earlier reduced reopens”), and feed the recent lessons to the step that makes the reply-or-escalate call, not to every step.
Notice the shape matches the lesson: record at decision time, wait for the real outcome, judge against a baseline, keep the lesson short, and hand it to the deciding step. The domain changed; the loop did not.
Flashcards
Section titled “Flashcards”Eight cards. Click any card to reveal the answer. Use the Print flashcards button to lay out the full set as one card per page for offline review.
Q. Why split recording a decision from judging it?
Because you cannot judge a decision before its outcome is known. Recording is instant and cheap; judging needs hindsight. So the system records now and reflects later.
Q. What does a 'pending' logbook entry mean?
The decision was made but its outcome is not yet known, so it has not been judged. Pending is the placeholder for hindsight, replaced once the reflection is written.
Q. When does reflection happen?
On the next run for the same company. At the start of that run, the system resolves earlier pending entries, by which point real time has passed and the outcome can be measured.
Q. How is the reflection grounded in something real?
It is fed the actual outcome over the following days (five by default), compared against a baseline of how a broad market slice moved (the framework uses the S&P 500). A baseline makes the test fairer than a raw up-or-down.
Q. Which model writes the reflection, and why?
The cheaper, faster model. Summarizing a known outcome is not the hard judgment, so it does not need the most capable model. Model tier follows difficulty.
Q. Which agent receives the accumulated memory?
Only the final gate, the portfolio manager, the one that makes the last call. Memory lands where the decision is committed, the same principle as spending the capable model there.
Q. What is the record-then-reflect pattern in one line?
Act, wait, look at what really happened, write one short lesson, and put that lesson where the next decision is made.
Q. Why ground reflection in a baseline, not just the raw outcome?
A baseline (what would have happened anyway) asks whether the work added anything, which is fairer than a raw up-or-down and harder to fool.