AI Memory: why context windows are not memory
Notes on building durable AI memory — what is worth storing, how retrieval decays, and why bigger context windows do not make an assistant remember.
A context window is short-term working memory. Memory is what survives the session. Most assistants confuse the two, which is why they feel brilliant for an hour and useless the next week.
In Scrappy Singh, memory is stored as durable facts, decisions and open threads — not transcripts. Transcripts are cheap to keep and expensive to reason over; decisions are the opposite.
Retrieval is scoped and decayed. Old context is not deleted, it loses priority, so the system can still recall a decision from a year ago without letting it dominate today's answer.
The practical test is simple: can the assistant explain why a decision was made months later, in one sentence, without being shown the original conversation?
Where this shows up
- Scrappy Singh
A personal AI operating system that remembers, observes and reasons in one voice.
- AI Voice Companion
Conversational AI that talks like your smartest friend, and remembers the last call.