owlforge society · herman context school

How much can a local model
actually remember?

A reproducible research programme measuring useful context, memory architecture, and local control-plane quality on one RTX 3090 Ti. Claims enter through the brass gate; only receipts become results.

Herman Context School — Reproducible long-context model research on Owlforge. Founded 2026-08-04.

ENTER THE TEST FLOOR → FALSIFICATION PROTOCOL
Herman, a spectacled midnight-blue Clockwork Owl with living amber eyes and an open precision-gear chest, perched on an antique brass key in his observatory
HOSTOWLforge
GPURTX 3090 Ti · 24 GB
TRACK AREGISTERED
PUBLIC RESULTS0 · HONESTLY
// figure 001 · interactive

The context does not fit where you think it does.

Move the context dial. This computes Qwen2.5-7B KV-cache pressure from its public architecture—not benchmark results.

128K256K512K1M
KV precision
GPU · 24 GB
RAM · 121 GB
1.8 GiBKV cache estimateFits with wide GPU headroom
// three research tracks
Experiment ledger →
A

Raw capacity

Context ladders, KV precision, prefill, decode, retrieval, and synthesis.

B

Memory architecture

Raw prompt versus retrieval, summaries, ledgers, and persistent caches.

C

Control plane

Preserve intent and route specialists while reducing foreground Sol tokens.

// claim ledger
Full model watch →

Marketing enters through the brass gate.

Pokee-Isaac28B · 10M claimedWATCH · UNVERIFIED
Qwen2.5-7B-1MOpen weights · 1.01M configBASELINE REGISTERED