Introducing Guru: The First Self-Evolving AI That Learns From Every Conversation
#1
by tejadabheja - opened
What is Guru?
Guru is a new AI architecture — no neural network, no GPU, no gradient descent. It stores knowledge as an editable graph and learns in real-time from every conversation.
Results
- 87% Exact Match after one round of corrections (from 1.8% cold start)
- 32ms average latency on direct lookups
- 54MB model (CSR) + 1.8GB knowledge base (LMDB)
- Runs on CPU, fits on a phone
What makes it different
- Learns from you: Every conversation improves the model. No retraining.
- Inspectable: Every answer traces back to specific knowledge.
- Editable: Delete a fact and it is gone. Instantly.
- Honest: Says "I don't know" when it doesn't. No hallucination.
- Safe by knowledge: Safety is taught as sentences, not hardcoded rules.
Try it
from brain_csr_adapter import BrainCSR
brain = BrainCSR(db_path=".")
brain.ask("What is the capital of France?")
# -> "The capital of France is Paris."
brain.correct("Who invented the telephone?", "Alexander Graham Bell invented the telephone.")
# Now it knows forever
Architecture
Co-occurrence graph + convergence loop (replaces attention) + WAL persistence + Q->A direct mapping with LRU cache. Two tiers: instant lookup for corrected questions, multi-hop graph reasoning for novel ones.
Links
Built by Tejas Phatak.
Feedback welcome!