forgemem 0.1.13

Persistent memory for AI agents — cross-session context that actually sticks

Forgemem

Persistent long-term memory for AI agents — cross-session context that actually sticks.

Forgemem mines your git history, session traces, and project notes to extract what actually happened — failures, successes, plans, and hard-won lessons. It stores them locally in SQLite and exposes them to AI agents via MCP so they start every session informed instead of blind.

---

The Problem

Every new agent session starts from zero. It doesn't know you already tried the Zod 4 schema approach and it broke. It doesn't know that wildcard CORS killed auth in production last quarter. It doesn't know which approach you abandoned and why.

Without Forgemem, agents repeat your mistakes. With it, they skip straight to what works.

---

Installation

pip install forgemem
forgemem init

forgemem init now requires a real TTY on first run so the user must choose an inference provider interactively. Agents cannot bypass that step with --yes or a piped session. After init completes, it auto-starts the MCP server; then restart your AI agent (Claude Code, Gemini CLI, or Codex) to pick up the new MCP connection.

---

How It Works

1. Mine — scans git repos and session notes, extracts failure/success/plan traces, saves them with an impact score
2. Store — all traces and distilled principles live locally at ~/.forgemem/forgemem_memory.db
3. Serve — an MCP server exposes the database to agents via the Model Context Protocol

Agents call retrieve_memories before starting any task. That one call surfaces the top matching failures and principles from your history — so they don't burn tokens re-learning what you already know.

---

Quick Start

# Install
pip install forgemem

# Initialize (interactive on first run)
forgemem init

# Optionally: also schedule background mining every hour
forgemem start --mine

# Restart your agent, then verify
forgemem status

---

Agent Support

Agent	Auto-detected	Skill file written
Claude Code	✓	~/.claude/skills/forgemem.md
Gemini CLI	✓	~/.gemini/forgemem-skill.md
OpenAI Codex	✓	~/.codex/forgemem-skill.json

forgemem init detects which agents are installed and writes the appropriate skill file automatically. Agents use retrieve_memories and save_trace via MCP — no extra configuration needed.

---

MCP Tools

Tool	Description
retrieve_memories	Search principles and traces by keyword before starting a task
save_trace	Save a failure, success, plan, or note during or after a session
mine_session	Agent-driven mining — no API key required, the agent is the LLM
distill_session	Condense raw traces into durable principles
forgemem_stats	Summary of DB contents and server health

---

CLI Reference

forgemem init                # Initialize DB, choose provider, register MCP, write skill files
forgemem start               # Start MCP server (background daemon on macOS)
forgemem start --mine        # Also install a scheduled mining agent (hourly)
forgemem stop                # Stop the daemon
forgemem status              # Show DB stats, server health, skill status
forgemem store "<text>"      # Save a memory trace manually
forgemem search "<query>"    # Search stored memories
forgemem mine                # Scan repos and session files for new learnings
forgemem distill             # Condense undistilled traces into principles
forgemem config              # Set inference provider (anthropic / ollama / gemini / …)
forgemem auth login          # Authenticate for managed inference (no BYOK needed)

---

Inference Providers

Forgemem uses an LLM for mining and distillation. Three options:

Provider	Setup	Cost
Forgemem managed	choose it in forgemem init, then run forgemem auth login	Free tier + paid plans
Ollama (local)	choose it in forgemem init	Free, fully private
BYOK (Anthropic / OpenAI / Gemini)	forgemem config <provider> --key <key>	Your API costs

forgemem init now requires the user to choose a provider interactively on first run. The mine_session and distill_session MCP tools still work with any subscription — no separate key needed since the agent itself is the LLM.

---

Platform Support

Platform	Daemon	Auto-start
macOS	LaunchAgent (launchctl)	On login
Linux	systemd user service (instructions printed)	Manual
Windows	Task Scheduler (command printed)	Manual

---

Licensing

Forgemem is Apache-2.0 for community use.

• Community — Apache-2.0, permissive, local-first, commercial-friendly
• Enterprise — hosted terms with SLA, SSO/SAML, audit logs, priority support, and private hosting
• Contributing — all contributors must sign the CLA by adding their name to CONTRIBUTORS.md in their first PR