Agent memory

LLM memory

Agent memory

STM allows the agent to maintain context and recall recent interactions within a single conversation or session

Short-term memory (STM)

Ref

- Model’s context window

--> Time passes and information begins to accumulate, it becomes increasingly difficult for the model to handle all the information effectively

STM cons

Ref

- Does not persist once the current task ends

Ex: solve a complex problem today, but if it isn’t specifically trained to remember this solution, it will have to start from scratch the next time

STM cons

Ref

- In-memory (Ex: LangGraph uses a checkpointer)

- In a temporary storage (Ex: Redis cache)

STM storage

LTM enables the agent to store and recall information across multiple sessions, allowing for personalization and learning over time

Long-term memory (LTM)

Ref

LTM storage

Real world example

Mermaid chard

flowchart TD

A[User sends new message] --> B[Append message to short-term buffer]

B --> C{Buffer size > k?}

C -- No --> D[Send short-term buffer + long-term summary to LLM]

C -- Yes --> E[Extract oldest messages beyond k]

E --> F[Summarize extracted messages with previous long-term summary]

F --> G[Update long-term summary]

G --> H[Remove extracted messages from short-term buffer]

H --> D[Send short-term buffer + updated long-term summary to LLM]

D --> I[LLM generates response]

I --> J[Append response to short-term buffer]

Github 

langgraph

Github demo code

Discussion?

LLM needs:

• Data (Knowledge)
• Dynamic sources (Pricing data, news, job postings...)
• Specialized content

Relationship

Ref

LLM benefits Web Scraping:

• Automatically define and parse elements, saving time and effort
• Define the extraction strategy and convert the extracted data into a structured format

Relationship

Ref

Demo

Ref

Crawl4ai

Discussion?