Useful bias manipulation re: LLM – the stochastic parrot speaks (gist.github.com)

In a recorded design session for an MCP (memory/coordination proxy) that guides an LLM during non-trivial coding tasks, the user and Claude converged on a simpler, more robust interface and a pragmatic view of LLM “self-models.” The key announcement: expose only two external tools—brain_query(intent: str, context: dict) -> Pattern and brain_submit(content: str, context: dict) -> SubmitResult—while keeping pattern creation, augmentation, routing and schema/version management inside the intelligent brain. The team updated modus-operandi.xml toward v2.0 to fix ambiguous file-vs-tool access, add concrete query examples, define error-driven workflows and submission criteria, and preserve “agency” language as a behavioral bias that improves outcomes. Significance: this reframes memory as an intelligent subsystem (not a dumb CRUD DB) and accepts that LLMs are “stochastic parrots” without intrinsic metacognition—so reliable behavior needs structural enforcement, not appeals to introspection. Practically, that means mandatory tool calls in system prompts, error-driven automatic queries, explicit keywords/triggers, and response formats that include query steps. The conversation also highlights trade-offs: useful “lies” about agency can bias token generation beneficially, but exposing internal ops or ambiguous triggers creates real contradictions. The result is a concrete, deployable pattern for integrating LLMs with an intelligent memory layer that reduces brittle, human-centric assumptions and improves repeatable performance.