Grafted Memory - GRM: A Routed, Tokenless K/V Memory Runtime

GRM (Graft Repository Memory) is a memory runtime for frozen language models that stores attention states rather than text: position-free K/V payloads ("grafts") harvested from live cache, routed by content, and re-seated into future contexts without re-reading. Four planes — VRAM as disposable mount arena (the model's live context window), RAM as authoritative repository, NVMe as write-ahead-logged durable storage, and a policy plane with evidence/fact separation, revision-instead-of-overwrite, and fidelity-gated consolidation. Gated results on consumer GPUs include 50-turn recall with the live context cleared every turn, recall through two generations of folded memory, 42-turn gates at flat device memory (~40 MB active), and warm-memory serving at 8x the prior production stack's job throughput. Prior work reuses cached K/V as immutable blobs; GRM operates on it — slicing grafts into children, classifying facts from evidence, superseding, folding, auditing. To our knowledge the conjunction — content-routed, re-seatable, durable, revisable attention-state memory on frozen models, across architectures — has no prior occupant.

Companion to "Selective Attention Is All You Need" (doi:10.5281/zenodo.21137970) and "Ghost Geometry" (doi:10.5281/zenodo.21138525), released together. Independent research, fully disclosed AI assistance. Text CC BY 4.0; implementation AGPL-3.0 (commercial licensing available).

Code: https://github.com/DragonShadows1978/GraftRepository (the GRM runtime: Python policy layer, C++ host mirror, gates and primers; opening alongside this release). Experiment platform: https://github.com/DragonShadows1978/AI-AtlasForge (MIT).