get_communication_history() renders raw dict instead of readable message text

[Harsh-617]

Describe the bug

get_communication_history() in EpisodicMemory, STLTMemory, and ShortTermMemory renders raw Python dictionaries instead of human-readable message strings when passed to the LLM prompt.

This happens because the speak_to tool stores message content as a nested dict:

# speak_to stores:
recipient.memory.add_to_memory(
    type="message",
    content={"message": "meet me at the north", "sender": 7, "recipients": [1, 2]},
)

Which results in entry.content["message"] being a dict, not a string. But all three get_communication_history() implementations format it directly into the f-string:

f"step {entry.step}: {entry.content['message']}\n\n"
# Actual output:
# step 5: {'message': 'meet me at the north', 'sender': 7, 'recipients': [1, 2]}

This garbled output is then injected into the ReAct reasoning prompt, degrading the LLM's ability to understand inter-agent communication.

Expected behavior

get_communication_history() should return a clean, readable string per message entry:

step 5: Agent 7 says: meet me at the north

To Reproduce

from unittest.mock import Mock
from mesa_llm.memory.st_memory import ShortTermMemory
from mesa_llm.memory.memory import MemoryEntry

agent = Mock()
agent.model.steps = 1
memory = ShortTermMemory(agent=agent, display=False)

# This is the structure speak_to produces
entry = MemoryEntry(
    content={"message": {"message": "meet me at the north", "sender": 7, "recipients": [1]}},
    step=5,
    agent=agent,
)
memory.short_term_memory.append(entry)
print(memory.get_communication_history())
# Current output:  step 5: {'message': 'meet me at the north', 'sender': 7, 'recipients': [1]}
# Expected output: step 5: Agent 7 says: meet me at the north

The same issue reproduces with STLTMemory and EpisodicMemory.

Additional context

• Affects all three memory implementations: ShortTermMemory, STLTMemory, EpisodicMemory
• The EpisodicMemory case is slightly different — _finalize_entry additionally wraps the content under the event-type key and appends an "importance" score, so entry.content["message"] becomes {"message": "...", "sender": 7, "importance": 3} — still a dict
• The bug silently corrupts the communication context window fed to the LLM on every reasoning step where agents have exchanged messages
• The existing test_get_communication_history test missed it because it constructed entries with a plain string "message": "Hello" rather than the nested dict structure speak_to actually produces