autogen-memorysync

MemorySync for Microsoft AutoGen (autogen-agentchat 0.4+): agents that remember users across sessions — without ever stalling a turn.

pip install autogen-memorysync

Quick start

from autogen_agentchat.agents import AssistantAgent
from autogen_memorysync import MemorySyncMemory

memory = MemorySyncMemory(
    api_key="ms_...",              # or MEMORYSYNC_API_KEY
    user_id="customer-42",         # required — who these memories belong to
    session_id="support-chat",     # tags the facts from this chat
)

agent = AssistantAgent("assistant", model_client=..., memory=[memory])

result = await agent.run(task=user_text)
# AutoGen never captures turns itself — capture the exchange in one call
# (the user's text is sent for fact extraction; the reply is not stored):
await memory.add_turn_pair(user_text, result.messages[-1].content)

update_context runs automatically before every model call: relevant memories are recalled and injected as a SystemMessage, and the retrieval surfaces to observers as a MemoryQueryEvent.

Why this one

Semantics worth knowing

• The turn is never stalled and never broken. Recall waits at most recall_timeout (default 1.2s); on timeout, outage, or quota exhaustion the agent simply answers without memories.
• user_id is required. No silent auto-generated UUID namespaces where stored memories can never be found again.
• User turns are sent to fact extraction; only durable facts are stored. add() sends a user turn to MemorySync as plain text with role: "user" under the autogen::<session> session scope. The server extracts the durable facts it contains and stores only those, tagged with session_id: "autogen::<session>" and your other metadata — the same shared user memories as every other MemorySync surface. Filler such as "ok" or "thanks" yields none, and the turn text itself is not stored. Deterministic idempotency seeds let the server recognise a retry of the same turn and not extract it twice.
• Assistant replies are not stored as memories. Content whose metadata["role"] is "assistant" (or "system", "tool", ...) is not sent, and add_turn_pair(user_text, assistant_text) sends only user_text. A MemoryContent without a role is the user's turn.
• By default clear() deletes the facts carrying this session's session_id and leaves the rest of the user's memories alone.
• Free-tier quota exhaustion is silent by design (adds accepted-without- storing, reads empty); evaluation keys surface strict 429s.
• create_memory_tools(memory) returns search_memory + save_memory FunctionTools for tool-equipped agents.

Configuration

Development

pip install -e . autogen-agentchat autogen-ext pytest pytest-asyncio "httpx>=0.25,<1"
python -m pytest tests -q      # 43 tests incl. a REAL AssistantAgent drive

Note: autogen-agentchat is in maintenance mode (Microsoft's successor is the Microsoft Agent Framework — a separate MemorySync adapter target). Maintenance mode means the Memory protocol this package implements is frozen and stable.

License

MIT