camel-memorysync

MemorySync memory for CAMEL agents.

Two pieces, use either or both:

• MemorySyncMemory — a drop-in AgentMemory for ChatAgent(memory=...): verbatim chat history plus semantic recall from MemorySync, injected as one scored context record under a hard latency budget. Zero client-side embeddings.
• MemorySyncStorage — a lossless BaseKeyValueStorage for ChatHistoryMemory that keeps the full history locally and sends the user's turns to MemorySync out of band, where only the durable facts in them are stored.

pip install camel-memorysync

Quickstart

from camel.agents import ChatAgent
from camel_memorysync import MemorySyncMemory

memory = MemorySyncMemory(
    user_id="customer-42",      # required — who these memories belong to
    session_id="support",       # conversation scope
)
agent = ChatAgent(system_message="You are a helpful travel assistant.", memory=memory)

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

The API key comes from the MEMORYSYNC_API_KEY environment variable (or api_key=...).

Why not the in-repo Mem0 storage?

CAMEL ships Mem0Storage in its repo. Verified against its source:

VectorDBMemory (CAMEL's own long-term option) needs client-side embeddings — OpenAIEmbedding() by default (extra key, cost, latency) plus a vector DB you operate. Recall here is one MemorySync query; the index lives server-side.

How recall enters the context

ScoreBasedContextCreator sorts records by timestamp. The recall block is one SYSTEM-role record with timestamp=0.0, so it lands right after the system prompt — your history is never reordered, rewritten, or role-swapped:

[system]  You are a helpful travel assistant.
[system]  Relevant long-term memories about this user:
          - Prefers window seats on long flights
[user]    which seat should I book for the Oslo flight?

Memories created by other MemorySync surfaces (LangChain, the CLI, voice agents, coding agents…) are recallable inside CAMEL too — that is the point of memory-as-a-service.

Storage-seam usage (ChatHistoryMemory)

from camel.memories import ChatHistoryMemory, ScoreBasedContextCreator
from camel.types import ModelType
from camel.utils import OpenAITokenCounter
from camel_memorysync import MemorySyncStorage

memory = ChatHistoryMemory(
    context_creator=ScoreBasedContextCreator(OpenAITokenCounter(ModelType.GPT_4O_MINI), 2048),
    storage=MemorySyncStorage(user_id="customer-42", session_id="support"),
)

save() keeps verbatim records locally (exact round-trip) and sends each user turn to MemorySync, which extracts the durable facts in it ("Prefers window seats") and stores only those; the turn text itself is not stored server-side, and filler is skipped. Assistant replies are not sent — they are not stored as memories — and system prompts and tool chatter never leave the process. A retried save is recognised server-side, so each turn is extracted once.

Deletion — designed against data loss

Configuration

Note on dependencies

camel-ai 0.2.x imports FastMCP from mcp.server, which mcp 2.0 moved — an unpinned install crashes at import camel.agents. This package pins mcp<2 until CAMEL supports mcp 2.

Tests

pip install -e . pytest
pytest tests -q   # 33 checks against the real camel-ai at latest

The suite drives a REAL ChatAgent turn via CAMEL's own StubModel and reproduces each in-repo Mem0Storage bug as a regression test.

License

MIT © MemorySync.