innernet memory

one memory for every ai tool, kept current by netti and answered from it. live, and measured on LongMemEval-S.

what it is

netti is the part of innernet that thinks. it folds what your tools send into one memory, keeps it current as things change, and hands any ai the few sections an ask needs, never the whole cabinet.

its recall is measured on LongMemEval-S, the long-memory benchmark most memory systems report: the full haystack, every question, graded with the benchmark’s own prompts. ask netti, the answers it gives you about your own memory, is judged live on a corpus of over a hundred asks.

recall is close to saturated across the field. what recall cannot see is whether a memory knows the person the facts belong to, which is where the person model starts.

it is live: use it at innernet.live

the evidence

LongMemEval-S, all 500 questions91.6% (458 of 500)packages/bench/BENCHMARKS.md, section 0 (run 2026-10-09, v5); results-aml-lmes-full500-haiku.json
held out, the 350 tuning never saw91.1% (319 of 350)packages/bench/BENCHMARKS.md, section 0 (run 2026-10-09, v5); results-aml-lmes-full500-haiku.json
its weakest kind, multi-session85.0% (113 of 133)packages/bench/BENCHMARKS.md, section 0 (run 2026-10-09, v5); results-aml-lmes-full500-haiku.json
reader and judgeclaude-haiku-5-5 · claude-sonnet-5-5, with LongMemEval's own judge promptspackages/bench/BENCHMARKS.md, section 0 (run 2026-10-09, v5); results-aml-lmes-full500-haiku.json

graded by Claude with LongMemEval's own prompts, not by gpt-4o, so it does not line up exactly with numbers graded another way
the other systems' scores are mostly self-reported, with different readers and judges
recall is close to saturated: seven systems report between 90 and 95 percent

papers and notes

nothing written up for this area yet. what it has is the evidence above.