AI Agent Memory: Search Past Sessions Without Starting Over

0:00AI engineers are making over 300 000 400 000 even 500 000 per year and this is part

0:083 of the AI systems design round that gets you these types of big tech jobs So if

0:14you are wanting to break into the AI industry stop and save this video So in part 1

0:21and part 2 we talked about an AI agent that calls tools appends the response to a conversation

0:28log The response if it's large is outside the conversation state so that it doesn't blow up the

0:35context window and that things are appended over and over and over again until a certain number of

0:41iterations or we get a tool call saying we're done and when the conversation gets too long we'll

0:46have compaction We'll summarize the conversation state and then we're going to continue In this part we're going

0:53to talk about memory and memory is actually pretty easy depending on your compaction architecture Now back in

0:59the day they always talked about RAG Retrieval Augmented Generation and I won't describe that in this video

1:07and it can be important but instead for memory here's a stronger solution Instead you're just going to

1:14essentially keep a read only log of everything that's ever happened without your AI agents and you're going

1:21to give the agent the tools to search past conversation state So it's going to do things like

1:29a semantic search and it's also going to do some sort of similarity search which yes can be

1:36RAG By implementing these search tools then not only can the current agent see what has done in

1:43the past but when you create new sessions those new sessions can see it too And this gives

1:50the agent some sort of long term memory So if you solved something before you can solve it

1:56again Now there are some things that you have to consider Sometimes things change over time Sometimes there

2:03are certain things that you do over and over and over again and that's when you want to

2:09have things like skills Also depending on the task you're going to want to have a different type

2:15of strategy If you're building an AI coding agent well how you coded something three months ago doesn't

2:21really matter that much at all does it You need an accurate representation of the code base right

2:27now so you might have some more complex or advanced strategies but you want to make sure that

2:34whatever you implement is specifically to target your specific goal But let's say you built your own call

2:41agent and you found out that it spent 1 000 trying to fix a super easy bug and

2:48you have no idea why you have no idea because you didn't implement observability or monitoring so like

2:55and follow for the final part