AG2 + MemorySync

Overview

Install ag2-memorysync and one attach call gives a classic AutoGen ConversableAgent long-term memory: every incoming user message is sent to MemorySync, which extracts the durable facts it contains, and is enriched with recalled context under a hard budget (default 1.2s) — through AG2’s own hook system, with zero extra code per turn. The agent’s own replies are not sent (assistant replies are not stored as memories); an outgoing message is sent only when its sender speaks for a person. The injected context feeds the LLM only; the original text is what is sent, so recalled memories can never re-enter memory. A process-wide dedup registry plus deterministic idempotency seeds avoid the double-store bug Zep documents in their own zep-ag2 adapter (two attached agents in one session store every utterance twice with conflicting roles — here each utterance is handled once, and an agent’s reply is never sent as the person’s words, by test). All async work rides one persistent background event loop that never touches the caller’s loop, so chats driven from inside asyncio work — zep-ag2 documents a deadlock caveat there. The package has zero framework dependencies: it duck-types against the agent, so it works with whichever classic distribution you installed and cannot version-conflict (zep-ag2 pins ag2<1).

Setup status and requirements

Supported version

ag2-memorysync 1.1.0 (PyPI)

Last setup review

2026-09-03

Permissions

A MemorySync API key with read and write scopes.

Limits

Targets the classic ConversableAgent framework (pip install autogen, Python 3.10+). The rewritten pip install ag2 v1 has no hook system yet — a future adapter target.

Capabilities

• One attach call — the whole integration
• Hard recall budget (default 1.2s) — the reply is never late
• User-turn capture with deterministic idempotency seeds; replies are not sent
• Multi-agent double-store bug (documented in zep-ag2) fixed by design, with a reproduction test
• Deadlock-free: one persistent bridge loop, never the caller’s (asyncio-driven chats pass, by test)
• Zero framework dependencies — no ag2<1 pin trap
• Injected context reaches the LLM but never the memory store
• search_memory + save_memory tools via register_memory_tools

Quick Start

from autogen import ConversableAgent
from ag2_memorysync import MemorySyncCapability

assistant = ConversableAgent("assistant", llm_config=...)

memory = MemorySyncCapability(api_key="ms_...",
                              user_id="customer-42",
                              session_id="support")
memory.add_to_agent(assistant)   # that’s the whole integration

Use Cases

Explore More