MEMORY AS A WASTING ASSET
SYSTEM MECHANISM EVIDENCE SIMULATOR PAPER
KAIKAKU · J. CHEN · 2026

A robot's memory physically wears out.

Every write to its onboard flash spends one of a few thousand erase cycles that never come back. This is the missing pricing layer for which memories are worth a cycle — and an interactive look at when it matters.

↓   SCROLL — OR JUMP TO THE  SIMULATOR
01

The argument, in five moves

A · THE ASSET

Memory is a wasting asset

TLC flash tolerates ~3,000 program/erase cycles per block. Each write spends one. The stock never refills.

non-renewable
B · THE PRICE

So price it like capital

One number — the rent on a single erase cycle, η — routes every memory: RAM, flash, cloud, or forget. The rule holds however value and write-rate correlate.

one price · all the work
C · THE TURN

The best memories may leave

When valuable memories are also the most rewritten, those are the ones to keep off local flash — they'd burn the chip out first. Value stops being monotone.

prop. 3 · χ > 0
D · THE NEGATIVE

Today, the lever is off

At datasheet prices the endurance budget never binds: η = 0. The learned controller doesn't beat simple price-based routing. The economics do the work.

honest negative
E · THE FUTURE

Priced for when it binds

Denser NAND, heavier writes, robots that revisit scenes for years — the budget tightens and η turns on. And it re-prices cleanly: when NAND prices spike in the 2025–26 supercycle, the equilibrium rent re-clears ~39% lower while the keep/evict boundary stays fixed.

regime-gated lever
02

The system

Any embodied-memory system answers three questions in sequence. Prior work solved the first. This paper supplies the other two — and the price that closes the loop.

FIG. 01 — RESEARCH PROGRAM
WHEN
write the memory at all
AURA · prior work
→
WHERE
which physical tier it lives in
this paper
→
WORTH
the price of an erase cycle · η
this paper
FIG. 02 — THE PRICED PLACEMENT LOOP
MEMORY-PRICE BAND · 2025–26 SUPERCYCLE AURA when to write prior work WEAR-AUGMENTED PLACEMENT INDEX value per byte, net of cash wear c·w and scarcity rent η·w route → argmax Πₓ RAM fast · volatile capacity rent ON-BOARD NVM (NAND) persistent · WEARS p_N·s + (c+η)·w the only line with rent CLOUD persistent · far p_C·s + latency tax FORGET write-off expected re-acquisition erases deplete the stock ENDURANCE STOCK · E_end ~3,000 P/E cycles · non-renewable the only constraint integrated over time η — budget binds
READAURA admits a memory (when); the wear-augmented index routes it to the tier with the highest net return Πˣ (where); writing to NVM depletes a non-renewable endurance stock, whose binding budget returns one scarcity price η that feeds back into the index (worth). One price closes the loop. After the paper's Fig. 2.
03

The mechanism

Placement is ordinary capital budgeting. Each tier return is a per-item income statement; only one line carries the wear, and that single line is what bends the optimum.

PER-BYTE PLACEMENT INDEX
I_i = λ_i (v_i+κ_i) / [ s_i (1−γe^(−δ_i)) ]
Value per byte of keeping memory i local: retrieval rate λ, value v, recompute cost κ, size s, staleness δ. Higher index, faster tier.
NET RETURN ON FLASH (NVM)
Π_i^N = V_i^fast − p_N s_i − (c_wear+η) w_i
Only the flash tier pays depreciation c·w plus the scarcity rent η·w. That extra term is the entire subject of the paper.
PERSIST-ON-FLASH RULE
B_i (value of locality) ≥ (c_wear+η) w_i
Keep i on local flash iff locality beats wear-plus-rent. η is the hurdle rate the scarce stock imposes; solved so the budget exactly binds.
WHY IT TURNS NON-MONOTONE
d/dv [ (c_wear+η) w̄(v) ] = (c_wear+η) χ
Locality is ~flat in value; wear cost rises in value exactly when χ > 0 (valuable memories get rewritten more). Then the most valuable items clear the bar last — persistence falls at high v. Proven; dormant at today's write rates.
04

What we measured

The whole non-monotone branch hinges on one empirical antecedent: is χ > 0 on real robots? Measured at a pre-specified gate before any controller was trained, the answer is not a law — it is a property of the deployment.

FIG. 03 — THE χ REGIME MATRIX
Deployment regime · backbone
sign
χ̂  (95% CI)
status
LIBERO-Long · SmolVLA-0.5Brecurrent, long-horizon manipulation
+
+1.0×10⁻³  [+0.38, +1.65]
pre-specified · holm-reject
LIBERO goal/obj/spatial · SmolVLAshorter-horizon suite
0
CI straddles zero
null
DROID · SmolVLAnon-recurrent teleoperation
−
−9×10⁻³  [−16, −4]
post-hoc · exploratory
OpenVLA-7B · LIBERO-Longcross-backbone agreement
?
ρ = 0.05 < 0.6 floor
uninterpretable
READThe sign of the value–write coupling χ tracks whether a robot keeps revisiting the same scenes. One confirmed, one null, one exploratory, one below the agreement floor. The simulator below lets you set χ by hand and watch the consequence.
05

The endurance economy, made playable

One synthetic population of memories, three linked views. Move the controls and watch the shadow price η re-clear, memories route across the tiers, and the persist curve bend. Everything below is solved live. The η shown is an illustrative toy quantity (sim units), not the paper's calibrated η — which is 0 today.

η price of one erase cycle χ do valuable memories get rewritten more? P/E program/erase — one flash write NVM on-board flash, the tier that wears
First — spend one block by hand
A robot's flash is built from blocks like this one. Each survives about 3,000 erase cycles, then it's dead. Saving a memory once costs one cycle; re-saving the same memory spends them fast — and nothing ever refills them.
3,000CYCLES LEFT · NON-RENEWABLE
A fresh block.
If this robot re-saves one memory to this block 4× a day, for 1 years…
Illustrative — one block, no wear-leveling. Real flash survives by spreading writes across many blocks; deciding which block and tier each memory lands on is exactly what this paper prices. Drag the orange numbers (or arrow-key them).
SCENARIO
WRITE LOAD0.90×
How write-heavy the deployment is.
ENDURANCE BUDGET1.15×
Cap on total erase cycles, as a multiple of unconstrained demand. Below 1.0× the budget binds.
VALUE↔WRITE χ+0.40
Do valuable memories get rewritten more (+), less (−), or independently (0)? The paper finds the sign is regime-dependent.
NAND PRICE REGIME
2025–26 supercycle band — raises cash wear. Higher NAND price re-clears the rent lower (∂η/∂price < 0): the direction behind the paper's −39%, shown here at toy scale.
POPULATION720 memories
Endurance shadow price
0 sim units
On local flash
—%
Evicted to cloud
—%
Forgotten
—%
Population — routed by the index
each glyph = one memory, binned by value (left low → right high), write-heaviest stacked on top
RAM NVM (flash) cloud valuable, evicted forget
Persist probability vs. value
how likely a memory stays on local flash, across the value range — bends down when χ>0 and η binds
P(keep local | v) monotone ref (η=0)