Week 1 agents made single-shot decisions — a call in, a label out, done. What is missing if you try to have a real back-and-forth with one of those agents?
The agent forgets. Every call is a fresh session, so it cannot reference what you told it on turn one when you are on turn three.
This week closes that gap. Five shapes of memory in five days — a numbered context string, keyword recall over a fact list, system_prompt injection, compressed summaries, and message_history for true multi-turn calls. By Friday your agents remember across turns.
And the memory lives in my Python code, not inside the model — I control what the agent sees on each call?
All memory is yours. A list, a dict, a string — whatever shape fits. The model only ever sees what you put in front of it. Week 2 teaches you the shapes.
build_context_string: numbered facts joined with newlinesrecall_facts: filter a fact list by keyword with any()agent_with_context: inject context via system_promptcompress_memory: summarize a long context into a short blurbstateful_response: multi-turn replies via message_historyGoal: by Friday you hand an agent the exact memory it needs — no more, no less.
Create a free account to get started. Paid plans unlock all tracks.
Week 1 agents made single-shot decisions — a call in, a label out, done. What is missing if you try to have a real back-and-forth with one of those agents?
The agent forgets. Every call is a fresh session, so it cannot reference what you told it on turn one when you are on turn three.
This week closes that gap. Five shapes of memory in five days — a numbered context string, keyword recall over a fact list, system_prompt injection, compressed summaries, and message_history for true multi-turn calls. By Friday your agents remember across turns.
And the memory lives in my Python code, not inside the model — I control what the agent sees on each call?
All memory is yours. A list, a dict, a string — whatever shape fits. The model only ever sees what you put in front of it. Week 2 teaches you the shapes.
build_context_string: numbered facts joined with newlinesrecall_facts: filter a fact list by keyword with any()agent_with_context: inject context via system_promptcompress_memory: summarize a long context into a short blurbstateful_response: multi-turn replies via message_historyGoal: by Friday you hand an agent the exact memory it needs — no more, no less.