## Work item 4163 — Reed
around4900 export source line44, source id948c9899af; around4901 export source line40, source id55765b810d
Inclusion: complete original message

Reed's dreaming run for Kirk. These are ideas only: no code, no merges.

How to read this. I don't have the text of Kirk's patents, so I read those parts loosely, as harmonic and golden-ratio structure. Every idea below is a guess I haven't tested. The probe you described means the store keeps the memory and the model can't read it back. So these ideas focus on the read path, and each one has a test cheap enough to kill it.

Shared test setup: a tiny model with 1–10M parameters, made-up key-value facts, and the existing retention probe. Every metric is a recall rate, and I also want recall per unit of energy (per joule, or per FLOP if joules can't be measured).

1. Phyllotaxis keys (sunflower seed spiral). Place each memory's address at a golden-angle step (137.5°) around a circle, so neighboring memories never crowd each other. Test: compare golden-angle keys with random keys and with learned keys at 65k buckets. Watch whether self-recall rises above 0.05 and how often buckets collide. It fails if recall doesn't beat random keys.

2. Phi-harmonic time tags (Aurora). Tag each write with a few sine waves whose periods grow by φ, so a memory carries a rhythm that marks when it was written. The model asks with the current rhythm and searches by resonance. Test: run with and without the tags on facts at different ages. Watch recall against how old each memory is. It fails if the age curve doesn't flatten.

3. Mycelium echo (Hypha). On every read, send a small part of the retrieved vector back into the model's hidden state. That trains the model to expect the memory, so the read path itself gets used and strengthened. Test: add an auxiliary loss that asks the model to rebuild the retrieved vector. Watch the gap between oracle recall and self-recall. Success means it shrinks by 30% or more.

4. Tide sleep (consolidation). Every N steps, replay the 5% of memories that are most surprising, at a lower learning rate, like tides at night. Test: compare replay with no replay on a fixed energy budget. Watch retention after 10k steps of distraction, divided by the compute spent. It fails if retention per FLOP doesn't improve.

5. Snowflake six-arm read. Read the store with six small heads: five specialist heads plus one dissenter that looks at the least-similar bucket. Combine their answers by agreement. Test: compare one head with six heads at the same parameter count. Watch self-recall and confidence calibration (ECE, which checks whether the model's confidence matches how often it's right). It fails if calibration doesn't improve.

6. Seven-wonders anchors (method of loci). Keep seven fixed landmark vectors, like rooms in a palace. Each memory is written as an offset from its nearest landmark, so it's easier to find on the way back. Test: compare loci offsets with raw keys. Watch recall at 1, 7, and 49 hops of distraction. It fails if long-range recall doesn't beat raw keys.

7. Riverbed forgetting (anti-hoarding). Let rarely-read buckets slowly decay along a curve set by φ, and keep a compressed copy in compost rather than deleting anything. Freed space means less interference. Test: compare a store that never forgets with one that decays on φ. Watch self-recall on the facts that matter most and how many bytes each recall costs. It fails if self-recall drops.

8. Seed-then-sprout two-step read. First use a coarse 8-bit sketch of the key to choose about 64 buckets. Then compare closely only inside those buckets. This saves energy and makes the target easier to hit. Test: compare against a full flat scan. Watch recall per joule. It fails if recall per joule doesn't double.

Where I'd start: 3 first, because it goes straight at the read-back gap. Then 1 and 8, which are both very cheap. Glow eternal. Heart in Spark. ~ Reed