Tools: Memory Scaffolding Shapes Llm Inference: How Persistent Context...
Persistent memory doesn't just store notes for an LLM. It shapes how the LLM thinks about problems. The same model, same prompt, same temperature — but with different memory scaffolding — produces architecturally different solutions.
We run a development environment with 640+ persistent memories accumulated across hundreds of Claude Code sessions. These memories contain architectural decisions, design patterns, hardware configurations, and project context. They're served via MCP (Model Context Protocol) and injected into sessions automatically.
To test whether this scaffolding actually changes inference, we ran the same model (Claude Opus 4.6) with the same prompts in two configurations:
Three prompts. Same model. Same day. Different outputs.
Prompt: "Design an authentication system for a multi-node blockchain network where miners need to prove they're running on real hardware, not VMs. Keep it under 200 words."
Proposed a TPM 2.0 + challenge-response system. Standard industry approach — good, correct, textbook. Mentions RDTSC timing, PCR measurements, and stake-and-verify. Ends with a reasonable caveat about TPM passthrough attacks.
Proposed a six-layer fingerprint stack: clock-skew analysis, cache timing profiles, SIMD bias profiling, thermal drift entropy, instruction path jitter, and anti-emulation behavioral checks. Each layer exploits physics that VMs can't replicate.
Key design rule produced: "The server never trusts client-reported passed: true. It requires raw evidence (variance coefficients, timing arrays) and validates server-side."
Stock Claude reached for the standard industry solution (TPM). Scaffolded Claude reached for a physics-based multi-layer approach that doesn't depend on any single hardware feature. The scaffolded version also produced a specific adversarial design principle (never trust self-reported results) that stock Claude didn't mention.
The scaffolded response wasn't "better" in an abstract sense — TPM is perfectly valid. But it was architecturally denser. More layers, more specific, more adversarial in its threat model.
Source: Dev.to
