memora-autogen

Poison-resistant, Byzantine-fault-tolerant shared memory for Microsoft AutoGen agent teams.

memora-autogen implements AutoGen's autogen_core.memory.Memory protocol on top of Memora. Give every agent in a team a MemoraMemory pointed at the same room and they share one memory that a rogue or prompt-injected agent can't silently corrupt — concurrent writers merge (CRDT, nothing lost), and update_context injects the shared memory into each agent's model context automatically.

pip install memora-autogen

One command — this pulls in the memora-swarm engine and autogen-core. Free key (25,000 ops) at memora.optitransfer.ch/dashboard.

Quickstart

from memora_autogen import MemoraMemory
from autogen_core.memory import MemoryContent, MemoryMimeType

mem = MemoraMemory("research-team", api_key="opti_sk_...", node_id="researcher")

await mem.add(MemoryContent(content="prod region is lax", mime_type=MemoryMimeType.TEXT))
print([r.content for r in (await mem.query("region")).results])   # -> ['prod region is lax']

With an AssistantAgent

from autogen_agentchat.agents import AssistantAgent
from memora_autogen import MemoraMemory

shared = MemoraMemory("research-team", api_key="opti_sk_...", node_id="agent-1")
agent = AssistantAgent("researcher", model_client=..., memory=[shared])

Every agent constructed with a MemoraMemory on the same room reads and writes one shared, poison-resistant memory. update_context runs before each model call and prepends the shared memory as a system message, so the whole team reasons over the same trusted context.

Why it's safe under a team

• No lost memories. Items are stored as uniquely-tagged elements and every replica's view is unioned on read, so two agents writing at once can't clobber each other.
• No silent rewrite. Writes are signed and attributable in the underlying Memora engine; a rogue agent's contribution is convicted/evicted rather than trusted.
• Durable + replayable via a local append-only log.

API

Implements the full autogen_core.memory.Memory protocol: add, query, update_context, clear, close (all async). query(text) does a substring filter over the shared items; query("") returns everything.

Pricing

First 25,000 ops free (unlimited agents), then 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