CAMEL-AI + MemorySync

Overview

Install camel-memorysync and pass MemorySyncMemory(user_id=..., session_id=...) to ChatAgent — verbatim chat history kept locally, plus semantic recall from MemorySync, injected as one scored SYSTEM-role context record that CAMEL’s ScoreBasedContextCreator sorts directly after the system prompt. Only the user’s turns are sent to MemorySync, which stores just the durable facts in them; assistant, system and tool records stay local. Your history is never reordered, rewritten, or role-swapped: the in-repo Mem0Storage returns extracted facts instead of your messages with every record hardcoded role=USER (breaking OpenAI’s alternation), and its clear() calls client.delete_users() — resetting a chat destroys the whole account. Here clear() resets the local window only, and remote deletion is explicit and loud: forget_session() and forget_user() delete the facts extracted from this session’s or all of the user’s CAMEL turns, so memories from other MemorySync surfaces survive. Recall runs under a hard 1.2s budget and fails open to history-only context, an anti-echo filter keeps the live conversation from coming back as “memories”, multimodal image payloads round-trip intact in the local history, and a deterministic seed lets the server recognise a retried save, so a turn is extracted once. Unlike VectorDBMemory there are zero client-side embeddings — no OpenAIEmbedding key, no vector database to operate.

Setup status and requirements

Supported version

camel-memorysync 1.1.0 (PyPI)

Last setup review

2026-10-01

Permissions

A MemorySync API key with read and write scopes.

Limits

Requires camel-ai 0.2.60+ (Python 3.10+).

Capabilities

• Byte-exact local history — roles, uuids, and images preserved
• Only user turns sent to MemorySync; replies stay local
• One scored recall record after the system prompt — history never touched
• Anti-echo filter — the live turn never returns as a “memory”
• clear() is local-only; forget_session()/forget_user() are explicit and loud
• Zero client-side embeddings — recall is one MemorySync query
• Hard 1.2s recall budget; fails open to history-only context
• Cross-surface recall — memories from LangChain, the CLI, voice agents work here
• Deterministic idempotency seeds — a retried save is extracted once

Quick Start

from camel.agents import ChatAgent
from camel_memorysync import MemorySyncMemory

memory = MemorySyncMemory(
    user_id="customer-42",      # required
    session_id="support",
)
agent = ChatAgent(
    system_message="You are a helpful travel assistant.",
    memory=memory,
)

agent.step("I always prefer window seats on long flights")
# a new session, days later:
agent.step("which seat should I book for the Oslo flight?")  # remembers

Use Cases

Explore More