AutoGen doesn't have a built-in execution cache. Every GroupChat, every ConversableAgent run starts fresh. If your multi-agent workflow runs similar tasks repeatedly — research pipelines, code review agents, scheduled reports — you're paying full LLM price every time.
Here's how to fix it without touching your AutoGen code.
The setup
bash
pip install mnemon-ai
import mnemon
mnemon.init()
# your existing AutoGen code — completely unchanged
import autogen
assistant = autogen.AssistantAgent(
name="assistant",
llm_config={"model": "gpt-4o", "api_key": "..."},
)
user_proxy = autogen.UserProxyAgent(
name="user_proxy",
human_input_mode="NEVER",
)
user_proxy.initiate_chat(
assistant,
message="Analyze Q2 sales data for Acme Corp and generate a summary report",
)
# second run with same or similar message: 2.66ms · 0 tokens · $0.00
Mnemon's MOTH layer patches AutoGen at startup. No agent changes, no conversation changes.
---
What gets cached
Every LLM call your agents make is intercepted. On repeat runs:
- Exact match — same message, instant response from cache
- Semantic match — "Analyze Q2 sales for Acme Corp" matches "Generate Q2 sales analysis for Acme" — same task, different phrasing
For multi-agent workflows where agents pass messages between each other, common sub-tasks (data parsing, formatting, summarization) hit the cache across different top-level goals.
---
For structured recurring workflows
If your AutoGen setup runs the same workflow repeatedly with varying inputs, use m.run() for segment-level caching:
import autogen, mnemon
m = mnemon.init()
def run_analysis(goal, inputs, context, capabilities, constraints):
user_proxy.initiate_chat(assistant, message=goal)
return user_proxy.last_message()["content"]
result = m.run(
goal="Q2 sales analysis for Acme Corp",
inputs={"quarter": "Q2", "client": "Acme Corp"},
generation_fn=run_analysis,
)
print(result["tokens_saved"]) # tokens saved on this run
print(result["cache_level"]) # "system1" | "system2" | "miss"
---
Numbers
┌─────────┬────────────┬────────────┐
│ │ First run │ Cached run │
├─────────┼────────────┼────────────┤
│ Tokens │ ~1,250 │ 0 │
├─────────┼────────────┼────────────┤
│ Latency │ ~20s │ 2.66ms │
├─────────┼────────────┼────────────┤
│ Cost │ full price │ $0.00 │
└─────────┴────────────┴────────────┘
At 80% hit rate on recurring workflows: 93% token reduction.
---
Install
pip install mnemon-ai # exact match only
pip install mnemon-ai[full] # + semantic matching (local, no API key)
import mnemon
mnemon.init()