# BEFORE (mem0)
from mem0 import Memory
m = Memory.from_config({"vector_store": {"provider": "qdrant", ...}})
m.add("fact", user_id="u1")
results = m.search("query", user_id="u1")

# AFTER (Memorus, ACE off — identical behavior, zero overhead)
from memorus import Memory
m = Memory(config={"vector_store": {"provider": "qdrant", ...}})
m.add("fact", user_id="u1")
results = m.search("query", user_id="u1")

# AFTER (Memorus, ACE on — full intelligent pipeline)
from memorus import Memory
m = Memory(config={
    "ace_enabled": True,
    "vector_store": {"provider": "qdrant", ...},
})
m.add(
    [{"role": "user", "content": "Always run pytest with -v flag"}],
    user_id="u1",
)
results = m.search("pytest verbose", user_id="u1")

from memorus import Memory

# Initialize with ACE enabled
m = Memory(config={"ace_enabled": True})

# Add knowledge from a conversation
result = m.add(
    [{"role": "user", "content": "Always run pytest with -v flag for verbose output"}],
    user_id="dev1",
)
# result includes: bullets_added, bullets_merged, bullets_skipped

# Search the knowledge base
results = m.search("pytest verbose", user_id="dev1")
for r in results.get("results", []):
    print(f"[{r['score']:.2f}] {r['memory']}")

# Without ACE, Memorus is a transparent proxy to mem0
m = Memory()
m.add("Some fact", user_id="u1")

Memory(config: dict | None = None)

config = {
    # --- ACE master switch ---
    "ace_enabled": True,

    # --- Reflector: distills conversations into knowledge bullets ---
    "reflector": {
        "mode": "hybrid",            # "hybrid" (default) | "rules" (0 LLM) | "llm"
        "min_score": 30.0,           # [0–100] minimum instructivity score to keep a bullet
        "max_content_length": 500,   # max chars per distilled bullet
        "max_code_lines": 3,         # max code block lines tolerated in a bullet
        # LLM settings (only used when mode = "llm" or "hybrid")
        "llm_model": "openai/gpt-4o-mini",  # any litellm-compatible model identifier
        "llm_api_base": None,        # custom API base URL (e.g. "https://api.deepseek.com")
        "llm_api_key": None,         # API key (falls back to env vars if None)
        "max_eval_tokens": 512,      # max tokens for LLM evaluation response
        "max_distill_tokens": 256,   # max tokens for LLM distillation response
        "llm_temperature": 0.1,      # low temperature for deterministic extraction
    },

    # --- Curator: deduplication and conflict detection ---
    "curator": {
        "similarity_threshold": 0.8,       # [0–1] cosine similarity threshold for merge
        "merge_strategy": "keep_best",     # "keep_best" | "merge_content"
        "conflict_detection": False,       # enable contradiction detection
        "conflict_min_similarity": 0.5,    # [0–1] lower bound for conflict window
        "conflict_max_similarity": 0.8,    # [0–1] upper bound for conflict window
    },

    # --- Decay: time-based forgetting ---
    "decay": {
        "half_life_days": 30.0,        # exponential decay half-life in days
        "boost_factor": 0.1,           # recall boost: weight *= (1 + boost × recall_count)
        "protection_days": 7,          # new memories immune to decay for N days
        "permanent_threshold": 15,     # recall_count >= this → permanent (weight=1.0)
        "archive_threshold": 0.02,     # weight below this → archive candidate
        "sweep_on_session_end": True,  # auto-sweep when session ends
    },

    # --- Retrieval: hybrid search tuning ---
    "retrieval": {
        "keyword_weight": 0.6,         # weight for exact + fuzzy + metadata layers
        "semantic_weight": 0.4,        # weight for vector search layer
        "recency_boost_days": 7,       # memories newer than N days get boosted
        "recency_boost_factor": 1.2,   # multiplier for recent memories
        "scope_boost": 1.3,            # multiplier for memories matching target scope
        "max_results": 5,              # max results returned after trimming
        "token_budget": 2000,          # max total tokens for LLM context
    },

    # --- Privacy: PII/secret sanitization ---
    "privacy": {
        "always_sanitize": False,      # sanitize even when ACE is off
        "sanitize_paths": True,        # redact OS user paths
        "custom_patterns": [           # additional regex patterns for redaction
            r"INTERNAL-\d{6}",
        ],
    },

    # --- Integration: agent behavior ---
    "integration": {
        "auto_recall": True,           # auto-recall during inference
        "auto_reflect": True,          # auto-reflect on tool results
        "sweep_on_exit": True,         # auto-decay on session exit
        "context_template": "xml",     # "xml" | "markdown" | "plain"
    },

    # --- Daemon: multi-process shared memory ---
    "daemon": {
        "enabled": False,              # enable background daemon
        "idle_timeout_seconds": 300,   # daemon shuts down after idle (5 min)
        "socket_path": None,           # IPC socket path (auto-resolved if None)
    },

    # --- mem0 backend config (passed through as-is) ---
    "vector_store": {
        "provider": "qdrant",
        "config": {"host": "localhost", "port": 6333},
    },
    "llm": {
        "provider": "openai",
        "config": {"model": "gpt-4o-mini"},
    },
    "embedder": {
        "provider": "openai",
        "config": {"model": "text-embedding-3-small"},
    },
}

m = Memory(config={
    "ace_enabled": True,
    "embedder": {
        "provider": "onnx",            # Use ONNX instead of OpenAI/etc.
        "config": {
            "model": "all-MiniLM-L6-v2",  # 384 dimensions (default)
        },
    },
})

config = {
    "privacy": {
        "custom_patterns": [
            r"INTERNAL-\d{6}",          # Company-specific IDs
            r"customer_[a-f0-9]{32}",   # Customer tokens
        ],
    },
}

# Global knowledge (shared across all users)
m.add("Always use UTC timestamps", scope="global", user_id="alice")

# Project-level knowledge
m.add("Use FastAPI for this project", scope="project:myapp", user_id="alice")

# User-specific (via mem0 parameter)
m.add("I prefer dark mode", user_id="alice")

# Agent-specific (via mem0 parameter)
m.add("Tool X requires --force flag", agent_id="tool_resolver")

# Search includes scope-matching boost (default +30%)
results = m.search("API patterns", user_id="alice", scope="project:myapp")
# Returns: project:myapp memories (boosted) + global memories

m = Memory(config={
    "ace_enabled": True,
    "daemon": {
        "enabled": True,
        "idle_timeout_seconds": 300,
    },
})
# All agents sharing this config use the same in-memory KB via IPC
# Falls back to direct mode if daemon is unavailable