Skip to content

The Open Source Memory Layer For Autonomous Agents

License

Notifications You must be signed in to change notification settings

galshubeli/Memary

ย 
ย 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 

Repository files navigation

memary_logo

LinkedIn Follow Documentation Demo PyPI Downloads Last Commit License

Manage Your Agent Memories

Agents promote human-type reasoning and are a great advancement towards building AGI and understanding ourselves as humans. Memory is a key component of how humans approach tasks and should be weighted the same when building AI agents. memary emulates human memory to advance these agents.

Quickstart ๐Ÿ

Install memary

  1. With pip:

Make sure you are running python version <= 3.11.9, then run

pip install memary
  1. Locally:

i. Create a virtual environment with the python version set as specified above

ii. Install python dependencies:

pip install -r requirements.txt

Specify Models Used

At the time of writing, memary assumes installation of local models and we currently support all models available through Ollama:

  • LLM running locally using Ollama (Llama 3 8B/40B as suggested defaults) OR gpt-3.5-turbo
  • Vision model running locally using Ollama (LLaVA as suggested default) OR gpt-4-vision-preview

memary will default to the locally run models unless explicitly specified. Additionally, memary allows developers to easily switch between downloaded models.

Run memary

Steps

  1. [Optional] If running models locally using Ollama, follow this the instructions in this repo.

  2. Ensure that a .env exists with any necessary credentials.

    .env
    OPENAI_API_KEY="YOUR_API_KEY"
    NEO4J_PW="YOUR_NEO4J_PW"
    NEO4J_URL="YOUR_NEO4J_URL"
    PERPLEXITY_API_KEY="YOUR_API_KEY"
    GOOGLEMAPS_API_KEY="YOUR_API_KEY"
    ALPHA_VANTAGE_API_KEY="YOUR_API_KEY"
    
  3. Fetch API credentials:

    API Info
  4. Update user persona which can be found in streamlit_app/data/user_persona.txt using the user persona template which can be found in streamlit_app/data/user_persona_template.txt. Instructions have been provided - replace the curly brackets with relevant information.

  5. [Optional] Update system persona, if needed, which can be found in streamlit_app/data/system_persona.txt.

  6. Run:

cd streamlit_app
streamlit run app.py

Basic Usage

from memary.agent.chat_agent import ChatAgent

system_persona_txt = "data/system_persona.txt"
user_persona_txt = "data/user_persona.txt"
past_chat_json = "data/past_chat.json"
memory_stream_json = "data/memory_stream.json"
entity_knowledge_store_json = "data/entity_knowledge_store.json"
chat_agent = ChatAgent(
    "Personal Agent",
    memory_stream_json,
    entity_knowledge_store_json,
    system_persona_txt,
    user_persona_txt,
    past_chat_json,
)

Pass in subset of ['search', 'vision', 'locate', 'stocks'] as include_from_defaults for different set of default tools upon initialization.

Adding Custom Tools

def multiply(a: int, b: int) -> int:
    """Multiply two integers and returns the result integer"""
    return a * b

chat_agent.add_tool({"multiply": multiply})

More information about creating custom tools for the LlamaIndex ReAct Agent can be found here.

Removing Custom Tools

chat_agent.remove_tool("multiply")

Memory Dashboard ๐Ÿง 

(coming soon)

Features Benefits
๐Ÿ—ฃ๏ธ Chat to Agent Memory Access certain memories
๐Ÿง  Analyze Agent Progress Track how agents develop their memories over time
โฎ๏ธ Rewind Executions Review agent memories to understand specific responses
๐Ÿง‘โ€๐Ÿง‘โ€๐Ÿง’โ€๐Ÿง’ Audience Preferences Understand audiences' best and most recent preferences
โœ๐Ÿป memaryParse Inject proprietary data into agent memory & combine parsers for advanced data ingestion
๐Ÿ—‚๏ธ memaryRetrieval Access memories & combine retrievers for advanced memory retrieval
๐Ÿงช Configure Agent Memory Search and combine memory databases
๐Ÿ› Playgrounds Specify models and tools used as well as benchmark different memory techniques
๐Ÿ” Stay Up To Date Receive notifications when agent memories have been added, updated or removed

Core Concepts ๐Ÿงช

The current structure of memary is detailed in the diagram below.

memary overview

At the time of writing, the above system design includes the routing agent, knoweldge graph and memory module are all integrated into the ChatAgent class located in the src/agent directory.

Raw source code for these components can also be found in their respective directories including benchmarks, notebooks, and updates.

Principles

memary integrates itself onto your existing agents with as little developer implementation as possible. We achieve this sticking to a few principles.

  • Auto-generated Memory

    • After initializing memary, agent memory automatically updates as the agent interacts. This type of generation allows us to capture all memories to easily display in your dashboard. Additionally, we allow the combination of databases with little or no code!
  • Memory Modules

    • Given a current state of the databases, memary tracks users' preferences which are displayed in your dashboard for analysis.
  • System Improvement

    • memary mimics how human memory evolves and learns over time. We will provide the rate of your agents improvement in your dashboard.
  • Rewind Memories

    • memary takes care of keeping track of all chats so you can rewind agent executions and access the agents memory at a certain period (coming soon).

Agent

routing agent

To provide developers, who don't have existing agents, access to memary we setup a simple agent implementation. We use the ReAct agent to plan and execute a query given the tools provided.

While we didn't emphasize equipping the agent with many tools, the search tool is crucial to retrieve information from the knowledge graph. This tool queries the knowledge graph for a response based on existing nodes and executes an external search if no related entities exist. Other default agent tools include computer vision powered by LLaVa and a location tool using geococder and google maps.

Note: In future version releases, the current ReAct agent (that was used for demo purposes) will be removed from the package so that memary can support any type of agents from any provider.

def external_query(self, query: str):
    messages_dict = [
        {"role": "system", "content": "Be precise and concise."},
        {"role": "user", "content": query},
    ]
    messages = [ChatMessage(**msg) for msg in messages_dict]
    external_response = self.query_llm.chat(messages)

    return str(external_response)
def search(self, query: str) -> str:
    response = self.query_engine.query(query)

    if response.metadata is None:
        return self.external_query(query)
    else:
        return response

Knowledge Graphs

KG diagram

Knowledge Graphs โ†” LLMs

  • memary uses a Neo4j graph database to store knoweldge.
  • Llama Index was used to add nodes into the graph store based on documents.
  • Perplexity (mistral-7b-instruct model) was used for external queries.

Knowledge Graph Use Cases

  • Inject the final agent responses into existing KGs.
  • memary uses a recursive retrieval approach to search the KG, which involves determining what the key entities are in the query, building a subgraph of those entities with a maximum depth of 2 away, and finally using that subgraph to build up the context.
  • When faced with multiple key entities in a query, memary uses multi-hop reasoning to join multiple subgraphs into a larger subgraph to search through.
  • These techniques reduce latency compared to searching the entire knowledge graph at once.
def query(self, query: str) -> str:
        # get the response from react agent
        response = self.routing_agent.chat(query)
        self.routing_agent.reset()
        # write response to file for KG writeback
        with open("data/external_response.txt", "w") as f:
            print(response, file=f)
        # write back to the KG
        self.write_back()
        return response
def check_KG(self, query: str) -> bool:
        """Check if the query is in the knowledge graph.

        Args:
            query (str): query to check in the knowledge graph

        Returns:
            bool: True if the query is in the knowledge graph, False otherwise
        """
        response = self.query_engine.query(query)

        if response.metadata is None:
            return False
        return generate_string(
            list(list(response.metadata.values())[0]["kg_rel_map"].keys())
        )

Memory Modules

Memory Module

The memory module comprises the Memory Stream and Entity Knowledge Store. The memory module was influenced by the design of K-LaMP proposed by Microsoft Research.

Memory Stream

The Memory Stream captures all entities inserted into the KG and their associated timestamps. This stream reflects the breadth of the users' knowledge, i.e., concepts users have had exposure to but no depth of exposure is inferred.

  • Timeline Analysis: Map out a timeline of interactions, highlighting moments of high engagement or shifts in topic focus. This helps in understanding the evolution of the user's interests over time.
def add_memory(self, entities):
        self.memory.extend([
            MemoryItem(str(entity),
                       datetime.now().replace(microsecond=0))
            for entity in entities
        ])
  • Extract Themes: Look for recurring themes or topics within the interactions. This thematic analysis can help anticipate user interests or questions even before they are explicitly stated.
def get_memory(self) -> list[MemoryItem]:
        return self.memory

Entity Knowledge Store

The Entity Knowledge Store tracks the frequency and recency of references to each entity stored in the memory stream. This knowledge store reflects users' depth of knowledge, i.e., concepts they are more familiar with than others.

  • Rank Entities by Relevance: Use both frequency and recency to rank entities. An entity frequently mentioned (high count) and referenced recently is likely of high importance, and the user is well aware of this concept.
def _select_top_entities(self):
        entity_knowledge_store = self.message.llm_message['knowledge_entity_store']
        entities = [entity.to_dict() for entity in entity_knowledge_store]
        entity_counts = [entity['count'] for entity in entities]
        top_indexes = np.argsort(entity_counts)[:TOP_ENTITIES]
        return [entities[index] for index in top_indexes]
  • Categorize Entities: Group entities into categories based on their nature or the context in which they're mentioned (e.g., technical terms, personal interests). This categorization aids in quickly accessing relevant information tailored to the user's inquiries.
def _convert_memory_to_knowledge_memory(
            self, memory_stream: list) -> list[KnowledgeMemoryItem]:
        """Converts memory from memory stream to entity knowledge store by grouping entities 

        Returns:
            knowledge_memory (list): list of KnowledgeMemoryItem
        """
        knowledge_memory = []

        entities = set([item.entity for item in memory_stream])
        for entity in entities:
            memory_dates = [
                item.date for item in memory_stream if item.entity == entity
            ]
            knowledge_memory.append(
                KnowledgeMemoryItem(entity, len(memory_dates),
                                    max(memory_dates)))
        return knowledge_memory
  • Highlight Changes Over Time: Identify any significant changes in the entities' ranking or categorization over time. A shift in the most frequently mentioned entities could indicate a change in the user's interests or knowledge.
  • Additional information on the memory module can be found here

Memory Compression

New Context Window

New_Context_Window

Note: We utilize the the key categorized entities and themes associated with users to tailor agent responses more closely to the user's current interests/preferences and knowledge level/expertise. The new context window is made up of the following:

  • Agent response
def get_routing_agent_response(self, query, return_entity=False):
        """Get response from the ReAct."""
        response = ""
        if self.debug:
            # writes ReAct agent steps to separate file and modifies format to be readable in .txt file
            with open("data/routing_response.txt", "w") as f:
                orig_stdout = sys.stdout
                sys.stdout = f
                response = str(self.query(query))
                sys.stdout.flush()
                sys.stdout = orig_stdout
            text = ""
            with open("data/routing_response.txt", "r") as f:
                text = f.read()

            plain = ansi_strip(text)
            with open("data/routing_response.txt", "w") as f:
                f.write(plain)
        else:
            response = str(self.query(query))

        if return_entity:
            # the query above already adds final response to KG so entities will be present in the KG
            return response, self.get_entity(self.query_engine.retrieve(query))
        return response
  • Most relevant entities
def get_entity(self, retrieve) -> list[str]:
        """retrieve is a list of QueryBundle objects.
        A retrieved QueryBundle object has a "node" attribute,
        which has a "metadata" attribute.

        example for "kg_rel_map":
        kg_rel_map = {
            'Harry': [['DREAMED_OF', 'Unknown relation'], ['FELL_HARD_ON', 'Concrete floor']],
            'Potter': [['WORE', 'Round glasses'], ['HAD', 'Dream']]
        }

        Args:
            retrieve (list[NodeWithScore]): list of NodeWithScore objects
        return:
            list[str]: list of string entities
        """

        entities = []
        kg_rel_map = retrieve[0].node.metadata["kg_rel_map"]
        for key, items in kg_rel_map.items():
            # key is the entity of question
            entities.append(key)
            # items is a list of [relationship, entity]
            entities.extend(item[1] for item in items)
            if len(entities) > MAX_ENTITIES_FROM_KG:
                break
        entities = list(set(entities))
        for exceptions in ENTITY_EXCEPTIONS:
            if exceptions in entities:
                entities.remove(exceptions)
        return entities
  • Chat history (summarized to avoid token overflow)
def _summarize_contexts(self, total_tokens: int):
        """Summarize the contexts.

        Args:
            total_tokens (int): total tokens in the response
        """
        messages = self.message.llm_message["messages"]

        # First two messages are system and user personas
        if len(messages) > 2 + NONEVICTION_LENGTH:
            messages = messages[2:-NONEVICTION_LENGTH]
            del self.message.llm_message["messages"][2:-NONEVICTION_LENGTH]
        else:
            messages = messages[2:]
            del self.message.llm_message["messages"][2:]

        message_contents = [message.to_dict()["content"] for message in messages]

        llm_message_chatgpt = {
            "model": self.model,
            "messages": [
                {
                    "role": "user",
                    "content": "Summarize these previous conversations into 50 words:"
                    + str(message_contents),
                }
            ],
        }
        response, _ = self._get_gpt_response(llm_message_chatgpt)
        content = "Summarized past conversation:" + response
        self._add_contexts_to_llm_message("assistant", content, index=2)
        logging.info(f"Contexts summarized successfully. \n summary: {response}")
        logging.info(f"Total tokens after eviction: {total_tokens*EVICTION_RATE}")

Future Features ๐Ÿ”œ

Speak to Your Agents Memory ๐Ÿ—ฃ๏ธ

memary's chat interface offers a portal to access agent memories, integrating capabilitiies such as searching, removing and viewing agent memories over specified periods and more all under one umbrella available in your dashboard.

Analyze Agent Progress ๐Ÿง 

Track how your agents develop their memory. We will provide access to relevant metrics to represent the growth of agent memories over time which will be available in your dashboard.

Track Memories โฎ๏ธ

memary breaks down agent memory for each response generated. A list of agent responses with their respective memories will be avilable in your dashbord. Human input (good/bad response) can help your systems improve.

Audience Preferences ๐Ÿง‘โ€๐Ÿง‘โ€๐Ÿง’โ€๐Ÿง’

Through our proprietary memory modules, we are able to infer audience preferences for certain time periods. Audiences' best and most recent preferences are continously updated and will be available in your dashboard.

memaryParse โœ๐Ÿป

Parse and clean your proprietry data before inserting into your agent memory. memary supports various file types including table and image extraction. Combine different parsers to form a parent parser with advanced capabilities. Also access templates for predefined database schemas and set of node relationships or define your own! This is all available in your dashboard.

memaryRetrieval ๐Ÿ—‚๏ธ

Use different techniques to retrieve agent memory. Also combine various retrievers to form a parent retriever with advanced capabilities. All avilable in your dashboard.

Customizable Memory ๐Ÿงช

memary deploys knowledge graphs to track agent actions. View, search and configure memory for your purposes. Join different memories together for improved retrieval and toggle between your favorite graph providers. All available in your dashboard.

Playgrounds ๐Ÿ›

  • Tool Playground: Simply define python functions and add it as one of your agent tools. View all available tools and remove any if necessary. Do this all in your dashboard!
  • Model Playground: Select specific models for tasks across memary to lower system LLM costs. All models deployed on HF will be avilable in your dashboard.
  • Benchmarking Playground: Easily run different memary configurations against each other to evaluate which memory options are more suitable for a specific task.

License

memary is released under the MIT License.

About

The Open Source Memory Layer For Autonomous Agents

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Jupyter Notebook 79.1%
  • Python 20.9%