total-agent-memory

total-agent-memory gives AI coding agents persistent memory of decisions, fixes and project conventions across sessions and repositories. It's for developers who don't want to keep explaining earlier work to Claude Code, Codex CLI, Cursor or another MCP client.

• Relevant recall: Text search, locally computed embeddings and a knowledge graph retrieve records related to the current task instead of loading the entire memory store.
• Changing facts and reusable lessons: Time-aware records preserve earlier values while newer facts replace outdated ones; procedural memory turns recurring error patterns into rules.
• Automatic capture: Claude Code hooks can save context in the background, and the native Hermes Agent plugin saves turns and recalls relevant records before the next turn.
• Shared team memory: A self-hosted server separates personal, department and company stores with access roles, authorship, revision history and a web interface for searching and editing records.

Memory lives in SQLite databases on your hardware. Default saves and searches work offline without an LLM call, API key or registration. Optional internal AI tasks can use Ollama, an OpenAI-compatible endpoint or Anthropic; remote providers receive the content sent for those tasks. These defaults apply to the memory service, not to the coding agent's own model requests.

The project is free and open source under MIT, runs on macOS, Windows and Linux, and offers Docker deployment. It requires Python 3.11 or newer and has no telemetry. A standalone CLI exposes memory to scripts, and the team dashboard can produce activity reports from stored records without an LLM.