LLM's Never Remember you

When ever you talk to an AI agent it the LLM never remember you because the llm doesn't have any memory about whome it's talking to and about what you are talking to.

LLM has a limited context window to make sure it remembers conversations we iteratively add the conversations to the llm. At a point the context window gets filled up it start forgetting the initial conversations which was given in the context window

If we iteratively do not add the conversations then the llm can't remember.

Memory makes an AI Agent more useful

The ability to

• Remember the past interactions
• Update information over time
• Maintaining long term and short term memory
• Forget outdated data
• Reasoning over historical data

Short term and Long term Memory

Short term memory is the storage where the entire messages are saved eg : Database examples :

• any situations that have happen between at some point of time
• random jokes happened at any point of time

Long term memory are the extract from the short term memory it only remembers key information

examples :

• name, personal info, key information is stored
• important dates, likes and dislikes

Episodic memory and semantic memory

In long term memory there are always 2 ways of storing any memory

Semantic memory

• it stores the straight forward facts to it
• these can be stored in the form of vector embedding because while retriving the data the semantic search helps to find relevent data
• similar kind of episodic data creates some facts about the conversation or user

Episodic memory

• it's a time series data based on incidents happening
• due to this episodic memory like i saw porsche, bmw, skoda it create a sematic fact that i like cars
• these are never deleted only add data into it

Why agent needs memory

Agent need memory is need to manage the sematic and episodic parts of a user and doing a post process on the memory so that we can make a stronger long term memory and update all the contradiction and delete unnessary episodes for a user

Memory makes the ai agent more personalised to the user

Memory Operation

• Write
• Update
• Delete memory is written into storages it might be normal db or vector stores and these memory is also updated as per the incoming events which are happening and the memory which is not further required also will be deleted

Mem0 and Knowledge Graphs

Mem0 acts as a memory layer for an ai agent it does all the above discussed funtionality it creates memory updates it deletes unwanted memory creates persistant long term memory does all the post process for stronger facts extracts from long term memory

Knowledge graphs are made using neo4j which is a graph data base it create nodes and links between the nodes giving it a relation. It works in bfs . while retrival if any related node comes then it takes the relevent data from graph and conciders it for the context of that user

Things an AI Agent should and shouldn't remember

Any agent should remember the fact and user preferences and any procedures which the ai agent should remember

Things which an agent shouldn't remember is any action which has completed or which are processing doesn't needed to be stored