Budgeted memory lifecycle — 20 Watts Was Enough

lusoris contributors

Purpose

Turn selective replay, consolidation, and forgetting into a costed decision problem. This note specifies a testable controller; it does not claim that its score is biologically implemented or optimal.

Memory items and actions

At maintenance cycle $t$ , the candidate memory set is $\mathcal{M}_t$ . Each item $i \in \mathcal{M}_t$ may receive one action

a_{i,t} \in \mathcal{A} = \{\text{defer},\text{replay},\text{merge},\text{externalize}, \text{weaken},\text{delete}\}.

The action is versioned. weaken and delete are invalid unless an external provenance record or explicit retention-policy exception exists.

Expected utility

For item $i$ and action $a$ , estimate

G_{i,a,t} = \widehat{\Delta L}_{i,a,t} + \lambda_R \widehat{\Delta R}_{i,a,t} + \lambda_S \widehat{\Delta S}_{i,a,t} - \lambda_D \widehat{D}_{i,a,t},

where:

$\widehat{\Delta L}_{i,a,t}$ is the predicted reduction in future task loss relative to defer (dimensionless loss units);
$\widehat{\Delta R}_{i,a,t}$ is the predicted reduction in a declared risk metric (risk units);
$\widehat{\Delta S}_{i,a,t}$ is the predicted improvement in schema-consistent transfer (dimensionless score units);
$\widehat{D}_{i,a,t}$ is predicted destructive-action harm, including rare memory loss and provenance failure (harm units);
$\lambda_R$ has units loss/risk;
$\lambda_S$ has units loss/score; and
$\lambda_D$ has units loss/harm.

The terms cannot be added before the conversion coefficients and evaluation sets are declared. Novelty, uncertainty, reward surprise, familiarity, and interference are features used to estimate these outcomes, not interchangeable units of utility.

Resource model

For each candidate action, predict:

$E_{i,a,t}$ — energy in joules;
$B_{i,a,t}$ — data moved in bytes;
$T_{i,a,t}$ — wall time in seconds; and
$W_{i,a,t}$ — slow-model optimizer updates (count).

The scheduler selects binary variables $x_{i,a,t} \in \{0,1\}$ :

\max_x \sum_{i \in \mathcal{M}_t} \sum_{a \in \mathcal{A}} x_{i,a,t} \left(G_{i,a,t} - \lambda_E E_{i,a,t}\right)

subject to

\sum_a x_{i,a,t} = 1 \quad \forall i,

\sum_{i,a} x_{i,a,t}E_{i,a,t} \le E_t^{\max}, \qquad \sum_{i,a} x_{i,a,t}B_{i,a,t} \le B_t^{\max},

\sum_{i,a} x_{i,a,t}T_{i,a,t} \le T_t^{\max}, \qquad \sum_{i,a} x_{i,a,t}W_{i,a,t} \le W_t^{\max}.

$\lambda_E$ has units loss/joule. The four budgets prevent a scheduler from appearing efficient by hiding memory traffic, time, or optimizer work behind a single energy proxy.

This is a multiple-choice multidimensional knapsack problem when predictions are fixed. The first experiment need not solve it exactly; greedy, learned, and standard replay priorities are compared under the same budgets.

The illustrative single-item price envelope shows how the upper admissible action can change with $\lambda_E$ . Its gains, costs, and selected actions are hypothetical and are not a fitted policy.

Reconsolidation gate

Retrieval produces a candidate branch rather than mutating the stored item in place. Let $m_{i,t}$ be a calibrated mismatch statistic and $\theta_t$ a gate:

z_{i,t} = \mathbb{1}[m_{i,t} > \theta_t].

If $z_{i,t}=0$ , the item remains read-only. If $z_{i,t}=1$ , an update branch is allowed but promotion still requires regression and provenance tests. Because the exact biological prediction-error boundary is disputed under C-040, $\theta_t$ is an experimental variable, not a biological constant.

Required ablations are:

always read-only;
always writable on retrieval;
fixed mismatch threshold;
calibrated risk-dependent threshold; and
explicit source-version change rather than inferred mismatch.

Schema-fit gate

Let $s_{i,t}$ be schema-fit estimated on a held-out structural probe and $q_{i,t}$ be shortcut risk. A candidate may receive a cheaper merge path only when

s_{i,t} \ge \tau_s \quad \text{and} \quad q_{i,t} \le \tau_q.

The thresholds $\tau_s$ and $\tau_q$ are calibrated on validation streams with both compatible items and deceptive near-matches. Training loss alone is not a schema-fit measure.

Measurement protocol

Report both realized and predicted quantities:

Quantity	Unit	Boundary
replay compute energy	J	accelerator or node, declared per run
memory traffic	byte	host–device and device memory reported separately
maintenance time	s	wall clock, including scheduler
optimizer work	update count	slow-model updates only
retention	task metric by item age	fixed evaluation stream
adaptation	events or seconds to threshold	after declared change point
obsolete intrusion	error count/rate	outdated association probes
destructive error	item count/rate	deleted state later shown necessary
provenance recovery	fraction	deleted items reconstructable from source

Calibration error between predicted $G$ and realized outcome is itself a core result. A scheduler that makes good-looking choices only after seeing future tests is invalid.

Falsification

The memory-lifecycle controller fails its first test if a conventional single-priority or reservoir baseline matches its quality–risk–energy frontier, if scheduling overhead consumes the saved replay budget, or if destructive errors exceed the declared safety bound.