cf memory

Manage the persistent AI memory system. Store, search, and recall project knowledge across Claude Code sessions. See Memory System for architecture details.

Subcommands

init

Interactive setup wizard. On first run, walks through each setting. On re-run (database exists), opens the config menu instead — same as cf memory config.

Wizard steps:

1. Search tier — auto, full (Tier 1), lite (Tier 2), or markdown (Tier 3)
2. Embedding provider — Transformers.js or Ollama
3. Auto-capture — auto-save session context on compaction
4. Auto-start daemon — start daemon when MCP server connects
5. Daemon timeout — idle timeout in minutes
6. Install deps & import — installs SQLite deps and imports existing memories
7. MCP setup — adds coding-friend-memory to .mcp.json

Each setting asks whether to save to global or local config.

Prerequisite for Tier 1: C++ compiler (Xcode CLT on macOS, build-essential on Linux).

cf memory init
# 🧠 Memory Setup
# Step 1/5: Search tier
# ? Memory search tier: full
# ...
# Memory initialized!

config

Interactive menu to modify memory settings. Same settings as the init wizard. Also accessible via cf config > Memory settings.

cf memory config
# 🧠 Memory Config
# ? Memory settings:
#   Tier [global] (auto)
#   Auto-capture [local] (true)
#   Embedding [global] (ollama/nomic-embed-text)
#   ...
#   Done

search

Searches memories from the terminal using Tier 3 substring matching. For full hybrid search (keyword + semantic), use the memory_search MCP tool inside Claude Code.

cf memory search "authentication flow"
# [0.85] JWT Authentication Pattern
#     id: features/jwt-auth | type: fact | matched: title, content

list

Without flags, lists individual memories in the current project. With --projects, lists all project databases sorted by size.

cf memory list --projects
#     5.7 MB    47 memories  912d2bfd5ad5  2026-03-16
#             /Users/you/project-a/docs/memory
#
#   213.1 KB     0 memories  0a1d7ea4a615  n/a
#             /Users/you/project-b/docs/memory
#
# Total: 5 project(s), 6.5 MB

rm

Removes project database(s). Markdown source files in docs/memory/ are not affected — only the SQLite index. Rebuild anytime with cf memory rebuild.

cf memory rm --project-id 912d2bfd5ad5   # Remove specific project
cf memory rm --all                        # Remove all databases
cf memory rm --prune                      # Remove orphaned projects

All commands ask for confirmation before deleting.

start-daemon / stop-daemon

Manages the Tier 2 daemon — a background server that keeps a MiniSearch index in RAM for fuzzy matching and BM25 ranking. Not needed if Tier 1 (SQLite) is enabled.

cf memory start-daemon
# Daemon started (PID 12345)

cf memory stop-daemon

The daemon runs detached and auto-rebuilds on file changes. By default, the daemon runs indefinitely until explicitly stopped with cf memory stop-daemon. If you configure an idle timeout and the daemon stops, the MCP server respawns it automatically on the next request (with up to 3 retry attempts before falling back to Tier 3).

rebuild

Re-reads all markdown files and rebuilds the search index from scratch. Run after manually editing memory files, switching embedding models, or pulling changes from git.

cf memory rebuild
# Rebuilt: 42 memories indexed.

status

Shows current tier, document count, daemon state, and MCP configuration.

cf memory status
# Tier: Tier 1 (SQLite + Hybrid)
# Memories: 42
# Daemon: stopped
# MCP: configured (local .mcp.json)
# Auto-capture: on

Configuration

In .coding-friend/config.json. Memory files are stored in {docsDir}/memory/ (default: docs/memory/).

{
  "memory": {
    "tier": "auto",
    "daemon": { "idleTimeout": 0 },
    "autoCapture": false,
    "autoStart": false,
    "embedding": {
      "provider": "transformers",
      "model": "Xenova/all-MiniLM-L6-v2"
    }
  }
}