Harness, Part 6: Memory
A memory made of one file: your agent remembers what matters - and still knows it after a restart.
What this is about
Since Part 2 your agent can remember a conversation - but only as long as the process lives. On restart everything is gone, because the history is short-term memory. Memory, the first extension from Block II, closes that gap - and the most honest form of it is strikingly plain: one file. It is loaded at start, the agent can add to it, and you can read it in your editor any time. No vector store, no database - for a personal agent the file is not the budget version, it is the right size. What bigger memory systems exist and when they pay off is told at the memory station.
A memory, though, raises a question no code can answer: what should your agent remember - and what expressly not? That is exactly what the interview is about this time.
At the end of this part, this works: you tell your agent a fact, restart it - and it still knows.
How the build works
As before: copy the prompt, hand it to your coding agent, answer its questions. From here on the series is Pro territory - the extensions require the complete Block I harness.
This section belongs to the Pro tier. You get access to full code, deeper explanations, and rabbit-hole material.
See tiers →It works when
A fact survives the restart: tell it, quit, start again, ask - it is still there. And the cross-check holds: whatever falls into your no-go zone, the agent declines. So far, though, it still needs you at the terminal. Part 7: Channels & Triggers changes that - your agent learns to run alone.