🧬 SoulAdapt

---

🎯 Why SoulAdapt?

Most AI companions either invent memories or treat every user exactly the same. SoulAdapt fixes both:

• 🎛️ Adapts how the AI talks: style, sensitive topics, interests
• 💎 Never invents: calibrates honesty to the real memory signal
• 🪶 Zero dependencies: pure Python standard library
• 🤝 Compatible, not dependent: works with SoulMemory or any object with .recall()

✨ Features

Feature	Description
observe()	Learn something about the user
learn_from()	Auto-extract observations from user text (ES/EN)
observations()	What the companion learned, strongest first
forget_observation()	Unlearn something
decide()	How to respond: style, avoid-list, interests, honesty, tone
prompt_context()	Ready-to-paste context string for LLM prompts
SincerityEngine	distance → confidence → assertive / hedged / admit (EN/ES)

📦 Installation

pip install souladapt

🚀 Quick Start

Standalone (no memory connected)

from souladapt import SoulAdapt

adapt = SoulAdapt("adapt.db")

adapt.observe("Prefiere respuestas cortas", category="style")
adapt.observe("Ruptura con Ana", category="sensitive")

d = adapt.decide("hola")
# → {'style': ['Prefiere respuestas cortas'],
#    'avoid': ['Ruptura con Ana'], 'interests': [],
#    'memories': [], 'confidence': None, 'honesty': 'neutral'}

Connected to SoulMemory

from soulmemory import SoulMemory   # optional extra: pip install souladapt[soulmemory]
from souladapt import SoulAdapt

mem = SoulMemory("memory.db")
adapt = SoulAdapt("adapt.db", memory=mem)

d = adapt.decide("¿qué sabes de Ana?")
# → honesty calibrated from real recall distances

💎 The three honesty levels

confidence ≥ 0.75 → assertive   "You had coffee with Ana."
0.45 – 0.75       → hedged      "If I remember correctly: ..."
< 0.45            → admit       "I don't have a clear memory..."

The AI never invents: if the memory is weak, it admits it.

🎭 Mood-based tone

When connected to a memory with emotional_timeline(), decide() reads the user's current mood and recommends a tone:

sadness / fear -> gentle (soft, warm) anger / disgust -> careful (calm, respectful) joy / surprise -> energetic (match the energy)

🤖 Auto-learning

adapt.learn_from('Prefiero respuestas cortas')   # -> style
adapt.learn_from("Me encantan los gatos")        # → interests
adapt.learn_from("No me hables de política")     # → sensitive

🤝 The contract (duck typing)

SoulAdapt never imports SoulMemory. Any object satisfies the contract if it has:

.recall(query, limit)  → list of dicts with "content" and "distance"

SoulMemory satisfies it out of the box. So does your own memory system.

️ Observation categories

style     → how to talk to the user ("short answers", "casual tone")
sensitive → topics to handle with care ("breakup with Ana")
interests → what they like ("cats", "gym")
general   → everything else

📚 API Reference

SoulAdapt(db_path="souladapt.db", memory=None)

Create an adapter. memory is any SoulMemory-like object (optional).

observe(content, category="general")

Register something learned about the user. Repeated observations get reinforced, not duplicated.

observations(category=None)

Get learned observations, strongest first.

forget_observation(observation_id)

Delete an observation: the companion unlearns it.

decide(query, limit=3)

Decide HOW the AI should respond. Returns style hints, topics to avoid, interests, evaluated memories, confidence and honesty level.

close()

Close the database connection.

🎬 Examples

python examples/demo_adapt.py   # standalone + connected to SoulMemory

🗺️ Roadmap

• Observations (reinforcement learning-lite)
• decide() adaptation layer
• SincerityEngine (3 honesty levels)
• Auto-learning (learn_from(), ES/EN)
• Mood-based tone adaptation
• Bilingual sincerity phrases
• prompt_context() for LLM integration
• Habit & routine detection
• Observation decay (unlearning over time)
• Multi-user support

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• SoulMemory — the memory layer this adapts to

---