Level 7 of 14 · about 3 weeks

Memory and context engineering

Decide what the model sees at each step, what it remembers across sessions and what lives outside the window.

Modules

1. 1

   What goes in the window

   choosing, trimming, ordering and fencing what the model sees at each step, and measuring whether the answer made it in

   Core 5 parts · about 8 hours

2. 2

   Short-term and long-term memory

   what an agent keeps between sessions, in which form, who is allowed to write it, and how to make it forget

   Core 5 parts · about 8 hours

3. 3

   Compaction and summarisation

   keeping a long-horizon agent under its token budget without losing the facts it still needs

   Core 6 parts · about 9 hours

4. 4

   Retrieval-based memory

   choosing which stored memories enter the window, and why "most similar" is not the same as "right"

   Core 5 parts · about 6 hours

5. 5

   Files and scratchpads

   using the file system as external memory, so the window holds handles and the data stays outside it

   Common 5 parts · about 5 hours

6. 6

   Context rot and long-context limits

   why a bigger window is not a better window, and how to measure where your model stops being reliable

   Core 3 parts · about 5 hours

7. 7

   Caching strategy

   ordering and editing the request so that most of it is read from the prompt cache, not recomputed

   Common 5 parts · about 4 hours

8. 8

   Isolating context in sub-agents

   giving each sub-task its own context window, and deciding what crosses back to the parent

   Common 5 parts · about 4 hours