I am sorry but the whole "biological memory" thing seems like marketing fluff on basic cache mechanisms.

You said it cuts token usage by 84% but isn't that typical for any typical chunked RAG system?

And why did you specifically chose to test against the LoMoCo dataset when there's a lot of issues with it and it being very easy to cheat?

Builder here, "Biological" is doing rhetorical work, agreed. Actual mechanism is exponential decay with category specific half lives, recall based reinforcement, and pruning below a strength threshold.

What separates it from a cache is that a cache evict by recency or LRU. This evicts by type, failures fade fast (environments change), strategies persist (what worked tends to keep working). And recall reinforces, so frequently used patterns survive even if old. A cache can't tell "I tried X once and it failed" from "I use X every week."

84% is vs storing everthing not vs well tuned chunked RAG. Smaller delta against good RAG, and I should have framed it that way. Where it beats rag is contradiction handling. RAG retrieves top k by similarity. If your old OAuth config and new OAuth config both score high, RAG returns both. Update + decay lets the stale one fall out.

LoCoMo, picked it because it's the closest thing to a standard for long context dialogue memory, and Mem0 reports on it. You're right it has issues. If you have a benchmark you trust more, genuinely interested.

Decay-as-eviction is just LRU, fair. Type-conditional half-life is worth defending, though.

A user's job and personality should be effectively permanent. Their stated intent for this week should fade in days. Their emotional state from a single message should be gone by tomorrow. Decay everything at one rate and you're back to LRU with the problems you're calling out.

The "biological" framing isn't really doing much work. Ebbinghaus is one curve and fine, but it's not where the leverage is. Type-conditional half-life is. Without that, this is a cache.

And a neural network is really just a composed, non-linear parameterized function that maps input vectors to output vectors. Sometimes metaphors or analogies do contribute something valuable.

I think it’s reasonable, a forgetting curve is intended to models a biological process.

https://en.wikipedia.org/wiki/Forgetting_curve