Within a company, semantic knowledge stores help agents answer questions better. While at Meta, we found that weaker LLMs with semantic layers consistently outperformed more powerful LLMs without them, both in terms of token cost and informational accuracy.
Grimoire.AI
Agents should learn together.
The only way to maintain a knowledge store for AI is with AI.
We know these things to be true
Semantic context is already winning.
Humans are already incapable of keeping up with the amount of AI code produced. A shockingly high percentage of code reviews have been outsourced to other AI.
Combining them, we reach a third truism
The only way to maintain a knowledge store for AI, is with AI.
The current paradigm
Sequestered agents protect platform lock-in.
OpenAI / Anthropic have insisted that agents within the same company must be sequestered and incapable of learning from each other. They would have you believe that this is because of privacy concerns such as the possibility of sharing personal information or API keys with another employee. This is a lie. They want to maintain two things:
They want each agent to spend more tokens repeating the mistakes of each other.
They want to maintain knowledge lock-in within their ecosystem. Anthropic wants all the important context needed to run your company to live within a bunch of stupid fucking markdown files on each user’s individual system, because it makes it harder to migrate knowledge over to another LLM provider. No matter how much Claude degrades, your company will be stuck with them.
At Grimoire, we reject this paradigm.
With one command
Save tokens. Get sharper answers. Keep moving.
- Save tokens
- Get more accurate answers
- With no additional effort