Déjà Vu Protocol (DVP)

「この仕事のノリ、前に対処した件と同じだわ」

Déjà Vu Protocol gives LLM agents semantic memory of past successes. When a new task is similar but not identical to a stored success pattern, it fires a déjà vu event — boosting dopamine and loading the proven GDC reasoning template.

Trigger Condition

if not exact_match and cosine_similarity(current_task, past_success) > 0.85:
    neurostate.dopamine += 30   # 「いける気がする」mode
    gdc.apply_template(past_success.gdc_shape)

Install

python3 -m venv .venv
.venv/bin/python -m pip install -e ".[dev]"

Python API

from dejavu import DejaVuEngine

engine = DejaVuEngine(threshold=0.85)

# Record a success
engine.record_success(
    task="fix the XSS vulnerability in login form",
    gdc_shape={"depth": 2, "nodes": 4, "branch": "feat/security"},
    neurostate={"joy": 70, "stress": 20},
)

# Check a new task
event = engine.check("sanitize user input in signup page")
print(event)
# [DVP] 🌀 Déjà vu fired! similarity=0.8823 dopamine+30 | 「この仕事のノリ、前に対処した件と同じだわ」

if event.fired:
    ns = engine.apply_neurostate_boost(event, {"dopamine": 50, "stress": 40})
    print(ns)  # {"dopamine": 80, "stress": 30}
    
    gdc_template = event.gdc_template  # load proven reasoning shape

How It Works

New task arrives
      ↓
Encode as vector (BoW / sentence-transformers)
      ↓
Compare against all stored success patterns
      ↓
exact_match?  → NO déjà vu (just repetition)
similarity > threshold?  → 🌀 DÉJÀ VU FIRES
      ↓
dopamine += 30  |  load GDC template  |  reduce stress

Embedder

Default: pure-Python bag-of-words (zero dependencies).
For production accuracy, swap to sentence-transformers:

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("all-MiniLM-L6-v2")

class STEmbedder:
    def encode(self, text: str) -> list[float]:
        return model.encode(text).tolist()

engine = DejaVuEngine(embedder=STEmbedder(), threshold=0.85)

Ecosystem

DVP integrates with the cognitive OS stack:

• NeuroState — receives the dopamine boost on déjà vu
• GDC — provides the reasoning template to load
• CPOS — kernel that can route tasks through DVP before dispatching

License

MIT