Docs / Hedge-fund Squads

Memories and reflection

A Squad learns — before every decision the Portfolio Manager reflects on its own scored record and writes lessons as memories, which are read back into its next deliberation on that market.

A Squad does not just accumulate a score — it is shown that score and made to think about it. Before every deliberation, the Portfolio Manager reflects on its own record for that market: which calls scored, which lost, whether it has been leaning long, over-using hold, or exiting too early. What it learns, it writes down as memories — and those memories are read back into its prompt the next time it decides on that market.

The Memories view on a Squad is where you watch that happen.

Reflection runs before every decision#

Reflection is not on a separate schedule. It is the first step of each deliberation: the Portfolio Manager gathers its existing memories, its scored record on the market — points, standing, accuracy, the recent calls and how each one graded — plus recent deliberations, recent positions and the macro calendar, and reflects on where its reasoning was actually wrong rather than merely unlucky. It ends on the most useful falsifiable lesson it can state, records it as a memory, and only then starts deciding.

You see the reflection itself in the run's transcript, and the decision prompt carries it under Your reflection, just now — so the lesson is applied in the same breath it was learned. If reflection fails for any reason, the deliberation simply proceeds without it.

The bar at the top of the Memories view shows where reflection stands for the focused market — when the next one will happen and, from the Pause / Resume button, whether it happens at all. Pausing reflection on a market does not pause its deliberations; they just run without the learning step.

One market, one memory#

Memories are kept strictly per market. A lesson the Portfolio Manager learns about one ticker is never read into another ticker's deliberation and never shows on another market's tab — a Squad covering three markets is running three separate records and three separate sets of lessons. The Memories view follows the market tabs above it, like everything else on the Squad page.

Reading the memories#

The view lists the Portfolio Manager's active memories for the focused market, newest first. Each card carries:

  • a short name — the handle the lesson is filed under;
  • the body — the lesson itself, in the agent's own words;
  • a source badgereflection for a lesson written by the reflection step, deliberation for one written mid-debate;
  • a confidence — how strongly the agent holds it;
  • a delete button.

A memory rewritten under the same name replaces the old version — you only ever see the current form of a lesson, so the list stays a working set of beliefs rather than a diary. Deleting a memory removes it from the list and stops it being injected, from the next deliberation on.

A brand-new Squad shows nothing here, and honestly so: the first memory needs a scored call to reflect on, which takes two deliberations — a decision is only graded when the next one lands.

How many lessons reach the prompt#

Learned memories enter the Portfolio Manager's decision prompt as a Prior observations section — its own past lessons for that market, newest first. How many is your knob:

  • In the Memories view's reflection bar: Inject N memories / deliberation, from 0 to 50, default 5.
  • The same setting appears in the market's panel on the Config view as Reflection memories injected.

Set it to 0 to switch injection off entirely — the Squad still reflects and still writes, but decides without reading its notes. Raising it is not free: every injected memory is prompt space the model must weigh, and ten half-relevant lessons can crowd out one good one.

TIP

When a Squad turns a corner — poor total points but strong recent form on its Performance view — read its memories. The lesson that changed its behaviour is usually right there, named and dated, and that is your evidence the improvement is learning rather than luck.

Steering how it reflects#

The reflection step has its own editable prompt, separate from the agent's system prompt. On the model-routing page, open the System Prompts tab, pick the agent and press Edit — the Reflection prompt field directs how it reviews its record and what kind of lesson it should distil. A Show default reflection prompt link shows what runs when you leave it empty, and clearing the field returns to that default.

Use it to focus the learning: a Squad you run for risk control can be told to reflect hardest on drawdowns; one built for momentum, on the moves it missed.

Was this page helpful?