Affective memory as the missing layer
Personaxis studies how *affect-weighted memory* can keep AI agents coherent, empathic, and accountable. Our work blends affective computing, cognitive science, and safety engineering to ensure that agents remember how interactions felt, not just what was said.
Affective signal intelligence
We engineer multi-modal pipelines that map raw text, prosody, and interaction telemetry into a shared affect space. Each interaction produces a vector (valence, arousal, dominance) plus uncertainty bounds. Focus areas:
- Lightweight prosodic conformers that infer mood from F0 variance, energy, and jitter—even on-device.
- Temporal alignment layers (differentiable dynamic time warp) so text and audio peaks synchronize.
- Behavioural priors that incorporate latency, retries, and opt-outs to calibrate confidence in real time.
Memory dynamics & personality coherence
Affective memories live alongside factual context. We weight each event by intensity so affect-rich experiences resist decay—mirroring how human memory works. Active investigations include:
- How mood parameters should evolve when sentiment oscillates rapidly (Bayesian smoothing + reinforcement signals).
- Constraints that keep personalities consistent—baseline trait vectors with limited daily drift to prevent “persona flips.”
- Explanation tooling that traces responses back to the most influential affective memories.
Socioaffective governance
Affective memory can amplify empathy—and risk. Our safeguards include:
- Consent-aware logging: every affect snapshot links to explicit consent state with one-click resets.
- Bias diagnostics: counterfactual replays expose disparities across demographics and trigger mitigations.
- Drift sentinels: affect variance thresholds force neutral fallbacks and operator review.
- Explainability hooks: governance consoles surface the memory chain and confidence metrics behind every response.
Current initiatives
- MVP affective SDK: affect-weighted retrieval, mood state machine, and regulator policies for de-escalation.
- Clinical & education pilots: measuring changes in trust, completion, and perceived care vs. baseline chatbots.
- Embodied agents: porting affect memory to expressive robotics to study physical co-regulation.
Want to collaborate? Reach out at davidq@personaxis.com or review the investor briefing for partnership details.