architecture
Serena's stack is not a model. It is a pipeline of symbolic operations on tokens, phonemes, and lifespan priors. Every layer reports its inputs, outputs, and the operation that connected them.
example trace · 9 nodes · 10 edges · trace_id 0x7af2
Input is tokenized, part-of-speech-tagged, and mapped to ARPAbet phonemes. No semantic embedding step. No vector store. Symbols stay symbols all the way down.
PMI lexicon assigns each token a tuple of emotional, valence, and intensity weights. A Kneser-Ney smoothed n-gram model gives conditional surprise. The combination is the emotional signature of the surface form.
Independent of the lexicon, a 7D affective vector is computed from non-semantic structural features — phoneme distributions, prosody proxies, syntactic patterns. Decoupling the 7D from the 12D guards against lexical bias dominating affect.
A six-dimensional trait vector parameterises the agent itself, not the input. STAB, OPEN, AGREE, ASSERT, SENSE, LEARN. Traits scale priors, modulate response shape, and persist across sessions.
Affect is age-shaped. Newborn blending, adolescent volatility, adult homeostasis. Every signal in L1–L3 is multiplied by a lifespan-scaled prior so the same input is read differently at different ages.
Short-term memory is held in an LMDB-backed adaptive resonance structure. Long-term memory uses minimal-perfect hashing for O(1) symbolic recall. Response generation is template-based and deterministic given the trace.