In Constant Itis’s small simulation, saved external memory appeared to restore a learned cue-to-action behavior on a fresh copy of a network—but the experiment did not use a reconstructed fruit-fly connectome or a biological brain. The author reports near-perfect task accuracy with external memory and similar performance before and after replacing the simulated network. Those results concern one simple task, one network setup and one fixed seed; they do not show that memories can be transferred between living brains.
What did “killing the brain” mean in this experiment?
It meant resetting or replacing a simulated network, not damaging an animal or a real brain. Despite the title’s use of “connectome,” Constant Itis says the tested substrate was a fixed recurrent network used as a stand-in—not a reconstructed fruit-fly connectome. The network was wired once and not rewired during the experiment.
As an Amazon Associate I earn from qualifying purchases.
The narrow question was whether experience stored outside that network could continue to influence behavior after the network’s learned state was erased, and whether a fresh equivalent network could use the saved state. The author describes this as preserving learned behavioral influence across destruction and replacement of the substrate that first experienced it.
How did the simulated task work?
The network received sensory inputs, processed them through a hidden middle layer and produced descending motor outputs. On each episode, it encountered one of six cues and had to select one of three actions. A fixed correct cue-to-action mapping was initially unknown; correct choices were rewarded and errors punished.
#1 Best Overall
- lucid fascinating mind brain
For the reported run, the author specifies 600 training episodes, 300 evaluation trials and one fixed random seed. Accuracy means the fraction of correct choices, with chance reported as 0.33. These are the author’s setup details, not an independent audit of the code or run.
What the external memory did
The associative memory was described as writing when salience crossed a threshold, completing patterns based on similarity, strengthening memories accessed together and allowing memories to decay. Salience was engineered from reward and novelty. Crucially, the memory could bias the network’s state but could not directly select an action; the network still produced the response.
How did the four configurations compare?
Constant Itis reports the following approximate accuracies for the cue-response task. All four figures come from the same described setup and single seed, so they should not be read as general performance estimates.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware match| Configuration | Reported accuracy | What the author reported |
|---|---|---|
| Fixed network alone | About 0.70 | The score reflected favorable initial preferences on that seed, not learning during training. |
| Fixed network plus plasticity | About 0.47 | Plasticity alone performed below the fixed network’s static score in this run. |
| Fixed network plus external memory | About 1.00 | The author describes this as the only condition that improved steadily on the task. |
| Plasticity plus external memory | About 0.88 | Adding plasticity reduced the reported score relative to external memory alone. |
The comparison supports a limited observation: in this particular simulation, the external-memory configuration performed best among the four tested conditions. It does not establish that external memory is generally superior to neural plasticity; only one architecture and one seed were tested.
Rank #3
What happened after the reset?
The author reports two reset comparisons. In the plasticity condition, accuracy fell from 0.44 before the reset to 0.20 after the network state and plastic weight changes were wiped. In the external-memory condition, the original scored 0.84, while a fresh equivalent substrate given only the saved memory scored 0.86.
| System | Before reset or replacement | After reset or replacement | What was retained |
|---|---|---|---|
| Plasticity | 0.44 | 0.20 | The changed plastic weights were erased along with state. |
| External memory | 0.84 | 0.86 | The saved associative memory was supplied to a fresh equivalent network. |
These are reported measurements from the described task. The author’s interpretation is that plasticity’s learned effect was stored in weights that the reset removed, whereas the associative state lived outside the substrate and could be reused. The result is evidence of transfer to a fresh equivalent stand-in in this setup—not transfer to a different kind of network, let alone a biological brain.
What the experiment does—and does not—show
What it shows within the simulation
- An external associative store could affect performance on a simple cue-to-action mapping without directly choosing the action.
- In the author’s reported run, the memory-equipped network scored about 1.00 in the main comparison, and a fresh equivalent network with saved memory scored 0.86 after replacement.
- Erasing plastic changes did not preserve the plasticity condition’s behavior in the reported reset comparison.
What it does not establish
- It is not a demonstration that a real fruit-fly connectome learns, or that a living brain can be reset and restored from an external memory store.
- The task was a small cue-response problem, not behavior in an embodied environment.
- Salience came from engineered reward and novelty signals, not measured or simulated real dopaminergic activity.
- The reported evidence covers one architecture and one seed; it does not establish performance across seeds, tasks or network designs.
- Transfer was tested to a fresh equivalent substrate, not to a different substrate class.
Constant Itis’s own characterization is appropriately narrow: “What I have shown is narrower: external associative state can preserve learned behavioral influence across destruction and replacement of the substrate that originally experienced it.” The distinction matters: saved state influencing a new equivalent simulation is an interesting proof of concept, but it is not biological memory transfer.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesCan readers reproduce the run?
The author points to a code repository and gives this command sequence:
git clone https://github.com/constant-itis/flymem && cd flymem && python3 flymem.pyThis is the stated route to run the code, not independent confirmation that the reported results reproduce. The figures above should therefore remain attributed to the author.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
