LlamaIndex + MemorySync

Overview

Install llamaindex-memorysync and pass MemorySyncMemory.from_defaults(user_id=...) to any agent: every user/assistant message persists to MemorySync the moment it happens — not when a 21k-token buffer eventually overflows, the silent-loss trap in block-only integrations — and recalled context injects through the framework’s own memory-block template. A genuine BaseRetriever makes memories RAG-queryable, and five structured tools give agents explicit memory.

Setup status and requirements

Supported version

llamaindex-memorysync 1.0.0 (PyPI)

Last setup review

2026-08-23

Permissions

A MemorySync API key with read and write scopes; the delete tool needs delete permission.

Limits

Requires llama-index-core >=0.13 <0.15 (Python 3.10+). Python only — LlamaIndex.TS has no memory-block architecture to integrate with yet.

Capabilities

• A real Memory subclass — every standard option and block composes
• Immediate persistence on every aput (the waterfall trap, fixed)
• from_defaults works; async-native end to end
• Recall through the native memory-block template with real partial truncation
• MemorySyncRetriever: memories in RetrieverQueryEngine and RetrieverTool
• Five never-raise agent tools with a read-only mode
• Serialization never leaks the API key

Quick Start

from llama_index.core.agent.workflow import FunctionAgent
from llamaindex_memorysync import MemorySyncMemory

memory = MemorySyncMemory.from_defaults(
    user_id="customer-7",
    session_id="thread-42",
)

agent = FunctionAgent(tools=[...], llm=llm)
response = await agent.run("Book my usual trip.", memory=memory)

Use Cases

Explore More