I Gave a Simulated Connectome External Memory and Killed the Brain
A developer attached an external Mycelium memory to a small recurrent network, proving that learned behavior survives a full substrate reset.
Stock photo for illustration only, not from the actual event
- A developer tested external memory integration on a small recurrent neural network stand-in.
- External memory successfully drove task accuracy up to approximately 1.00.
- Saved memories survived a full substrate destruction and reinstated behavior in a fresh brain.
- The experiment uses a simplified model and single seed, not a real fruit fly connectome.
A software developer has published an experimental concept exploring whether a simulated connectome can utilize external memory. The investigation asks a narrow question: can externally stored experiences survive a complete substrate reset and reinstate behavior in a brain that never originally learned them? The setup bypasses the 166,700-neuron connectome for now, utilizing a small fixed recurrent network as a stand-in.
The architecture incorporates Mycelium mechanics featuring salience-gated writes, similarity-based pattern completion, co-access strengthening, and data decay. The experiment enforces a strict core rule: the memory system cannot choose actions directly; it can only bias the internal neural state from which those actions naturally emerge.
Stock photo for illustration only, not from the actual event
The evaluation relies on a cue-response gauntlet task where the agent maps specific cues to correct actions. Accuracy ranges from 0 to 1, with random chance sitting at 0.33. The fixed network happened to start at roughly 0.70 due to random initialization favoring certain mappings, but a static brain can never improve on its own without plasticity or memory storage.
Testing across four conditions showed that isolated plasticity resulted in noisy performance near 0.47. In contrast, external memory was the only mechanism driving steady improvement up to ~1.00. Combining both yielded ~0.88, indicating that the two systems do not cleanly add together and naive plasticity creates slight interference.
"The external memory survived. The original system scored 0.84. A fresh substrate given only its saved memory scored 0.86."
The Developer
In the headline destruction test, the substrate was completely wiped back to a blank slate. The original system scored 0.84, while a brand-new substrate equipped solely with the saved external memory scored 0.86. This demonstrated that the learned behavioral influence did not survive inside the biological tissue, but rather survived outside of it.
This experiment draws a fascinating parallel to stateless computing architectures, where state data is decoupled from the processing compute unit. While the author notes significant caveats—such as using a hand-wired stand-in network and single-seed testing rather than a true dopaminergic fruit fly connectome—it opens up compelling theoretical discussions on memory persistence and data portability across artificial substrates.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment