memora-langchain 0.1.0

Poison-resistant, Byzantine-fault-tolerant shared chat memory for LangChain agent swarms — one prompt-injected agent can't corrupt the shared history.

memora-langchain

Poison-resistant, Byzantine-fault-tolerant shared chat memory for LangChain agent swarms.

When several LangChain agents share one conversation memory, that memory is the weakest link — one prompt-injected or malfunctioning agent can overwrite what every other agent then reads. memora-langchain backs LangChain's message history with Memora, a CRDT + Byzantine- tolerant shared store: concurrent writers merge without losing turns, and no single agent can silently rewrite the shared history.

pip install memora-langchain

That's it — this pulls in the memora-swarm engine and langchain-core for you. Get a free API key (25,000 ops free) at memora.optitransfer.ch/dashboard.

Quickstart

from memora_langchain import MemoraChatMessageHistory

history = MemoraChatMessageHistory(
    "support-crew",              # room: every agent using this room shares one history
    api_key="opti_sk_...",
    node_id="triage-agent",      # unique per agent
)

history.add_user_message("customer can't log in")
history.add_ai_message("checking their account status…")

for m in history.messages:       # merged view across the whole swarm
    print(type(m).__name__, m.content)

Use it with RunnableWithMessageHistory

from langchain_core.runnables.history import RunnableWithMessageHistory

chain_with_memory = RunnableWithMessageHistory(
    your_chain,
    lambda session_id: MemoraChatMessageHistory(session_id, api_key="opti_sk_...", node_id="agent"),
    input_messages_key="input",
    history_messages_key="history",
)

MemoraChatMessageHistory is a standard langchain_core.chat_history.BaseChatMessageHistory, so it works anywhere LangChain accepts a chat history — including as the shared store behind a LangGraph node.

Why it's safe under a swarm

• No lost turns. The history is stored as uniquely-tagged message elements and every replica's view is unioned on read, so two agents writing at once can't clobber each other — both sets of turns survive.
• No silent rewrite. Every write is signed and attributable to a node in the underlying Memora engine; a rogue agent's contribution is convicted/evicted rather than trusted on faith.
• Durable + replayable. State survives restarts via a local append-only log.

Pricing

The first 25,000 ops are free (unlimited agents). After that it's metered per semantic op — see memora.optitransfer.ch. Reads and keepalive are never billed.

Links

• Engine & docs — https://memora.optitransfer.ch
• Source & other integrations — https://github.com/mgillr/memora-swarm
• Core package — memora-swarm

Apache-2.0 · a product of optitransfer.ch