default.toml

# ICM Configuration
#
# Location (platform-specific):
#   macOS:   ~/Library/Application Support/dev.icm.icm/config.toml
#   Linux:   ~/.config/icm/config.toml
#   Windows: C:\Users\<user>\AppData\Roaming\icm\icm\config\config.toml
# Or override with: ICM_CONFIG=/path/to/config.toml

[store]
# SQLite database path (default: platform data dir).
# Uncomment to use a custom location:
# path = "/custom/path/to/memories.db"

[memory]
default_importance = "medium"

# Decay rate per day (0.95 = loses 5% per day)
decay_rate = 0.95

# Weight threshold for automatic pruning
prune_threshold = 0.1

[embeddings]
# Set to false to disable embeddings entirely (no model download, keyword search only)
# enabled = false

# Embedding model (fastembed model_code). Default: multilingual-e5-small (384d, 100+ languages)
# Other options:
#   "BAAI/bge-small-en-v1.5"              — 384d, English-only (fastest)
#   "Xenova/bge-small-en-v1.5"            — 384d, quantized English-only
#   "intfloat/multilingual-e5-base"       — 768d, multilingual (more accurate, slower)
#   "intfloat/multilingual-e5-large"      — 1024d, multilingual (best accuracy, largest)
#   "Xenova/all-MiniLM-L12-v2"            — 384d, English
#   "jinaai/jina-embeddings-v2-base-code" — 768d, code-optimized
model = "intfloat/multilingual-e5-base"

[extraction]
# Layer 0: rule-based fact extraction (zero LLM cost)
enabled = true

# Minimum keyword score to keep a fact
min_score = 3.0

# Maximum facts per extraction pass
max_facts = 10

# Extract every N PostToolUse hook fires (i.e. every N tool calls in a
# Claude Code / Gemini / Codex / Copilot session). Lower = more
# aggressive (more memories, more noise); higher = quieter. The
# `--every` CLI flag on `icm hook post` overrides this.
extract_every = 3

# Store the raw tool output as a low-importance memory when the
# rule-based extractor produces no facts. Useful as a coarse safety
# net; set to `false` if you only want signal-rich extractions.
store_raw = true

[recall]
# Layer 2: context injection before sessions
enabled = true

# Maximum memories to inject
limit = 15

[wakeup]
# SessionStart hook (Layer 1): wake-up pack of critical/high-importance
# memories injected at session start.

# Token budget for the pack (~4 chars/token). Increase for richer context;
# decrease if your tool's system prompt is already tight.
max_tokens = 500

# Include preference / identity memories regardless of project filter so
# user-wide settings (style, language, conventions) leak through into
# every project's wake-up.
include_preferences = true

[mcp]
# MCP server transport: "stdio"
transport = "stdio"

# Custom instructions appended to MCP server description
# instructions = "Always recall before starting work"