A Memory That Never Sleeps

Building Minds — what a firm's AI memory should attend to, keep, and bring back

A Memory That Never Sleeps

Until August, the decision memory I had been building for investment firms kept every word of every document it had ever read. Nothing faded.

That design feels safe. It is not. The remedy is not deletion.

A memory that retains everything at equal weight has never decided what matters. The next person who searches decides for it. Most days that is enough. It fails in three ways, and each failure costs judgment the firm has already paid for.

Similar decisions blur. Two add-on acquisitions with the same structure and different outcomes answer the same query; nothing says which lesson applies. Popularity begins to look like truth. Material cited often rises in the ranking, and if summaries draw on what ranks highly, repetition passes for corroboration. Old warnings stay buried. Evidence that didn't matter when it arrived, but matters a great deal after the thesis changes, has no path back to the surface.

Simpler fixes exist, and they may be enough: better ranking, strong provenance, date filters. I do not yet know whether a more deliberate design beats a well-tuned single store. That is the honest starting point for what follows.

What the brain does at night

The brain faces the same problem on a larger scale. Much of its answer arrives during sleep.

It runs two learning systems. The hippocampus captures experience quickly, one episode at a time. The neocortex learns slowly, extracting what holds across many episodes. McClelland, McNaughton, and O'Reilly argued in 1995 that the split is necessary: a single system fast enough to learn from one experience would overwrite what it had learned before.

Most of what the fast system captures does not last. Frey and Morris found in 1997 that a weak change at a synapse became durable only when strong stimulation elsewhere on the same neurons triggered protein synthesis within a window of less than three hours. They proposed a tag that marks which changes are eligible to be kept. Selection also looks forward. In a 2011 study by Wilhelm and colleagues, sleep improved memory only for people who had been told they would be tested later.

At night, the hippocampus replays the day. In 1994, Wilson and McNaughton found that rat place cells that fired together during exploration fired together more often in the sleep that followed. Two years later, Skaggs and McNaughton showed that the firing order was preserved as well. In 2009, Girardeau and colleagues tested whether the replay matters: suppress those bursts after training, and rats learn a maze more slowly.

Sleep also takes things away. Tononi and Cirelli's synaptic homeostasis hypothesis holds that waking strengthens synapses across the cortex and sleep scales them back down, keeping the differences between them while restoring room to learn. In mice, de Vivo and colleagues found synaptic contacts about 18 percent smaller in animals that had slept than in animals that had been awake, with the largest synapses spared. Richards and Frankland argue that this kind of forgetting is a feature: it keeps outdated detail from interfering with flexible decisions.

None of this proves that software should work the same way. It offers design hypotheses, and each one translates into an operation.

Capture fast; integrate slowly. New evidence counts the moment it arrives. Changes to what the firm believes go through a slower step. Score attention, and let the score fade unless something renews it. Integrate by comparison: before proposing a change to a belief, pull the prior evidence that supports it and the prior evidence that conflicts with it. Scale priority down as well as up, or everything drifts to the top. Let what falls away leave the default view, not the record.

One departure from the brain is deliberate. Recalling a human memory can change it. Nader, Schafe, and LeDoux showed in 2000 that reactivated fear memories in rats needed new protein synthesis to re-stabilize. A firm's memory should never work that way. Using a memory can change how prominent it is. It must never change the evidence behind it.

A firm cannot simply forget

Here the analogy breaks, and the break matters. Firms carry audit, legal, and regulatory obligations. The document that justified a decision cannot vanish because nobody opened it for a year.

So the design separates two things a brain never has to separate. The record keeps everything, unaltered, with its provenance. Active memory is what search and reasoning consult by default. For a firm, forgetting means demotion from active memory — never destruction of the record.

Five things a single store blurs

Once the record and active memory are separate, five properties that usually hide inside one relevance score come apart.

Attention priority: what gets reviewed, integrated, or shown first. Evidence confidence: how strongly a claim is supported. Temporal validity: when a claim applies, and whether it has been superseded. Memory status: whether something is recent, consolidated, or archived. Retention protection: what must stay recoverable, and under what conditions.

Holding them apart yields four rules.

Use is not evidence. Retrieval and citation can raise priority; they never raise confidence. Authority counts from day one. Audited financials carry evidentiary weight the moment they arrive, before anyone has worked through what they imply. Staleness is a validity question, not doubt. A figure reported as of a date is no less supported a year later; it is only less current. Protection is not prominence. A decision record can be guaranteed never to be destroyed without being guaranteed the top of every search.

The first rule is the one I needed most. The system has an option, built but never switched on, that would let approvals, upheld challenges, and recent citations nudge a fact's confidence upward. This work argues for keeping the first two and dropping the citations.

A Memory That Never Sleeps