This repository holds enhanced Agents, built for the Microsoft AutoGen Framework. Debuting with a MemoryEnabledAgent, with improvements in context/token control, portability, and Plug-and-Play functionality.
The Memory Enabled Agent (MEA) inherits from the AutoGen AssistantAgent class, with changes and features that endow the agent with a persistant, autonomous, and token-stable method of conversation, memory storage, and memory retrieval. This work is inspired by, and extends from, lessons and experiences from developing AutoGen_MemoryManager and AutoGen_IterativeCoding, as well as discussions within the AutoGen community. This work represents a continuation in the efforts to develop a more generally-powerful AI Agent. Feedback and Collaboration is strongly welcomed and encouraged.
The MEA was designed to provide an easy-to-implement option for using an Agent with dynamic and self-arranged memory within the AutoGen framework. The implemented system was inspired by literature on human memory systems, and hence, includes a Working Memory (CC, Chat-Context), Short-Term Memory (STM, via exChat-Context), and a Long-Term Memory (LTM, via exSTM-Context). The STM stays affixed to the top of the CC and is excluded from FILO concerns. The CC is a fixed number of messages which, when exceeded, uses a compression ratio (CR_1) to remove some of the oldest messages (FILO) and have them be summarized into the STM. The STM also has a fixed number of memories which, when exceeded, uses a CR (CR_2) to remove some of the oldest memories and integrate them into the LTM. Both summary events are executed by the internal MemoryEnabledAgent_Manager, referred to as the Memory Manager Agent, or MMA.
The Chat-Context (CC) exceeding the limit is what drives all memory storage related functions. When the CC exceeds the limit, a Compression Ratio (CR) is applied onto the messages such that:
trim_index = len(messages)*CompressionRatio
lost_messages = messages[:trim_index]
remaining_messages = messages[trim_index:]The trim_index splits the chat history per the ratio, with the oldest section (lost_messages) being sent to the Memory Manager Agent (MMA) for processing and appending to the Short-Term Memory (STM), with the newest section (remaining_messages) remaining as the new CC.
The MMA, when presented with the chat section, is prompted:
Conversation Section to Summarize:\n{lost_messages}\n\n Please make a function call to append_to_short_term_memory and pass in the key points you can extract from the above conversation section.
The MMA will then call append_to_short_term_memory and pass in a list of memories meant to capture any significance from the lost_messages.
The STM behaves in a similar fashion to the CC, in that it has a limit which, when exceeded, drives further memory storage functions. When the STM exceeds the limit, a CR (independent of CC CR) is applied such that:
trim_index = len(STM)*CompressionRatio
lost_memories = STM[:trim_index]
remaining_memoriess = STM[trim_index:]The trim_index splits the STM per the ratio, with the oldest section (lost_memories) being sent to the MMA for processing and incorporation into the LTM, and the newest section (remaining_memories) remaining in the STM.
The LTM is the final destination for all memories. There are no checks for max length. The MMA is supposed to maintain a minimal list, but some tuning may be required to achieve optimal performance.
Memories are retrieved from the LTM via a function call from the Memory Enabled Agent (MEA). The MEA will pass a hint to the MMA, that is to describe what information is being requested. The MMA is to return an answer to the query - not a copy-paste of the existing memory. This is intentional to allow the blurring/combining of seperate memories, if relevant.
The following default memory parameters have been used:
# Max number of short term memories before initiating compression to long
DEFAULT_SHORT_TERM_MEMORY_LIMIT = 10
# Proportion to cut short term memory off (0.9 drops 9 out of 10 memories after exceeding STM limit, 0.1 drops 1 out of 10 memories after exceeding STM limit)
DEFAULT_COMPRESSION_RATIO_STM = 0.8
# Max convo length before the end starts falling off, in number of messages. User and AI both count, so minimum is 2.
DEFAULT_MAX_CONVO_LENGTH = 10
# Proportion to cut chat off (0.9 drops 9 out of 10 chats after exceeding limit, 0.1 drops 1 out of 10 chats after exceeding limit)
DEFAULT_COMPRESSION_RATIO_CHAT = 0.8One should carefully consider the implications of changing these values. For any flow, there may be a balance to achieving performance with minimal tokens/requests, but it is somewhat case-by-case. There have been many precautions taken to prevent token overflow, in an attempt to lower costs. In reality, this agent is likely to have a higher minimum-token-useage. It is when discussions become long, or have the potential to become long, that the MEA may present an attractive solution.
The author assumes you already have AutoGen installed and running on your system. If you would like to simply use the system as a conversational agent, ensure main.py has your API configs filled, and is in the same directory as EnhancedAgents.py. You can run it from command line via python main.py.
You can converse with the agent normally, just use auto-reply for memory lookup requests.
The documentation in the current state is incomplete, and represents a work in progress.
- Fix 'User' and the actual user name not being consistent.
- Better prompts for better factoid generation/storage. Sometimes it still does sentences or other copy/paste.
- Automation of CC, STM, and LTM function calls, after ensuring minimal/zero waste.
- Code TODOs.
- Create a self reflection cycle to allow memory compression/reformating outside of CC driven demands. Make it accessible to the MEA.
- Create a context-switch cycle, to allow the STM to be reformated from CC/STM/LTM such that it is best prepared for the next task.
- Create a system to allow MEA to adjust STM and LTM-lookup for specific agent - either through more files or better control of entity names in STM and LTM entries.
- Devise a methodology and rationale for splitting LTM into related sections, to allow a larger total LTM but with less cost per LTM lookup if in a properly related section.
- Test and refine methods to achieve optimal performance and efficiency when using the agent in other activities outside of user chat (coding groups, media creation, cross-group ideation, etc...)