Build emotion-aware agents that appraise experience, preserve what mattered, revise their identity safely, and bring generative intelligence directly into SQL workflows.
The emotion-memory architecture models experience through identity alignment, outcome evidence, expectation violation, recurrence, salience, and retention. Success, failure, wound, and trauma become operational states rather than decorative labels.
The rest of the kit keeps that research practical: Gemini agents, tools, RAG, structured output, multi-agent patterns, MCP-style integrations, dataframe analysis, machine-learning automation, and local SQL AI enrichment.
A deterministic pipeline converts event and outcome evidence into appraisal, affective state, identity pressure, retention, retrieval, and policy evidence. Corrections rebuild derived state from the active journal so invalid evidence cannot haunt future decisions.
Identity alignment, valence, expectation violation, salience, recurrence, success confidence, failure confidence, and evidence quality are computed explicitly before an affective state is produced.
Appraisal produces repeat pressure, avoidance pressure, identity-update pressure, retention floors, expectation updates, and operational outcome regions.
Ingest, correct, delete, and expire operations remain auditable. Derived state is rebuilt from active evidence.
Candidate memories pass relevance gates before emotional retention and trajectory deduplication influence the final selection.
Experimental primitives explore event-tree attention, state encoding, temporal-difference updates, and memory-shaped action selection.
Parse SQL pseudo-functions, execute the relational query locally, enrich the resulting rows with Gemini in parallel batches, and merge the generated columns back into a dataframe or replacement table.
-- Add an AI-generated customer insight to every selected row
SELECT
customer_id,
feedback,
ai_generate(
prompt='Write one sharp customer-experience insight',
model='gemini-2.5-flash'
) AS customer_insight
FROM feedback;-- Summarize long support transcripts inside a normal SQL workflow
SELECT
ticket_id,
transcript,
ai_summarize(
prompt='Summarize the issue in twenty words or fewer',
model='gemini-2.5-flash'
) AS summary
FROM support_tickets;-- Classify each row into one constrained business taxonomy
SELECT
article_id,
body,
ai_classify(
prompt='Classify the business topic',
labels='Retail | Supply Chain | Pricing | Marketing'
) AS topic
FROM articles;-- Create a new table containing extracted retrieval signals
CREATE OR REPLACE TABLE enriched_browse AS SELECT
session_id,
page_text,
ai_extract(
prompt='Extract high-signal shopping-intent keywords',
count=5
) AS retrieval_keywords
FROM browse_logs;The project moves from general agent foundations to specialized workflows without forcing every use case into one orchestration abstraction.
Register Python functions, generate Gemini tool schemas, and let the model invoke capabilities when needed.
Use JSON agents when downstream systems need parsed objects rather than prose.
Chunk, embed, store, retrieve, and answer from private knowledge with grounded context.
Profile dataframes, train baseline models, detect leakage, compare metrics, and automate modelling workflows.
Run orchestrator, parallel, supervisor, conversation, and debate-style coordination patterns.
Expose and consume tools using lightweight Model Context Protocol style server and client components.
Install from PyPI, set your Gemini API key, and start with the simple agent core. The emotion-memory architecture and SQL AI modules are available when the system needs to become more ambitious.
# Install
pip install agentic-ai-kit
# Configure
export GEMINI_API_KEY="your-key-here"
# Build
from agentic_ai.agents import BaseAgent
agent = BaseAgent(
name="Sage",
sys_prompt="Be useful and concise.",
)
print(agent.think("What matters here?"))