Governed memory for AutoGen multi-agent conversations — audit chains, poisoning defense

[lanbass869-cell]

Governed memory for AutoGen multi-agent conversations — audit chains, poisoning defense

Hi AutoGen community! I built MemTether — a governed memory layer specifically designed for multi-agent scenarios where you need to know who wrote what, when it was true, and whether it was tampered with.

Why not just use built-in memory?

AutoGen's built-in memory is great for session context. MemTether adds governance:

• Every fact carries a mandatory source identity (which agent wrote it)
• Corrections supersede — never delete. Query what the system believed at any past date
• Cross-source contradictions queue for human adjudication (agents propose, humans decide)
• Hash-chained audit log — tampering is detectable
• OWASP LLM01/02/06 poisoning defense built in

Integration

REST API or MCP server. Works alongside AutoGen's native memory:

# After each agent response, record it
import memtether
memtether.remember(f"{agent_name}: {response_summary}", source=f"autogen-{agent_name}")

# Before the next turn, recall governed context
results = memtether.search("relevant query", limit=5)

GitHub | PyPI

Would love feedback from anyone running multi-agent crews where memory trust matters!