Real runs, graded objectively. Pick a benchmark to see the conclusion first — then the evidence. The same discipline applies to every claim MemoryLayer makes: verify before you believe it.
shadcn-ui/ui · 59% cheaper overall · 6 questions
Same model, same questions, same repo — only the tools differ. fs_list / fs_grep / fs_read vs a single recall(explore) call over 859 pre-ingested symbols. Tokens from usage.input_tokens, both sides prompt-cached.
Token comparison
| Type | Question | fs tools | recall(explore) | Δ cost | Correct? (judge) |
|---|---|---|---|---|---|
| literal | What error message does the useSidebar hook throw when called outside a SidebarProvider, and what is SidebarProvider's default value for its `defaultOpen` prop? | 2 turns1,530 tok$0.0064 | 2 turns5,118 tok$0.0154 | +141% | fs ✓ · recall(explore) ✓ |
| literal | How is the Dialog component structured and what primitive does it wrap? | 3 turns10,907 tok$0.0337 | 2 turns5,070 tok$0.0152 | −55% | fs ✓ · recall(explore) ✓ |
| literal | Which file defines the cn class-name utility and what does it combine? | 2 turns3,529 tok$0.0112 | 2 turns5,089 tok$0.0153 | +37% | fs ✓ · recall(explore) ✓ |
| semantic | Which components in this repo support keyboard shortcuts or display key bindings? | 3.3 turns19,836 tok$0.0606 | 2.3 turns7,477 tok$0.0224 | −63% | fs ✓ · recall(explore) ✓ |
| semantic | How does this codebase express form validation and error messages? | 4.7 turns32,266 tok$0.0983 | 2 turns4,815 tok$0.0144 | −85% | fs ✓ · recall(explore) ✓ |
| semantic | What components are available for visualizing tabular data or charts? | 3 turns9,188 tok$0.0285 | 2 turns4,723 tok$0.0142 | −50% | fs ✓ · recall(explore) ✓ |
Model: claude-sonnet-5. Tokens from Anthropic API usage.input_tokens (exact, real agentic loop). Repo: shadcn-ui/ui (859 symbols). $/Q at Opus list pricing ($3/M input).
With recall(explore) as the entry point and prompt caching enabled on the client, the realistic average is 59% cheaper than the same Opus loop with fs tools (N=3average). The savings over multi-tool MemoryLayer are unconditional; the rest comes from caching, which is one line in any Anthropic SDK integration. And that's before counting cross-session memory, where fs tools have nothing to offer.
Add persistent memory in 5 minutes.
One namespace — every agent, chat, and teammate.