memgentic

Universal AI memory layer — zero-effort knowledge capture across every AI tool you use. Source-aware memory with semantic search, filtering, and knowledge graphs.

memgentic is the core engine. Install it to get the CLI, the MCP server, the ingestion pipeline, and the storage layer.

pip install memgentic
# or with the REST API:
pip install 'memgentic-api'
# or with native Rust acceleration:
pip install 'memgentic[native]'

What it does

Every AI conversation is ephemeral — knowledge is lost when the session ends. Worse, it's siloed: what Claude knows, ChatGPT doesn't. memgentic captures knowledge automatically from every AI tool you use and makes it available everywhere via MCP.

• Local-first. Data stays on your machine. Works offline.
• Source-aware. Every memory carries provenance: which tool, which session, which timestamp.
• Semantic search. Qwen3-Embedding-0.6B via Ollama, 768-dim, Apache 2.0.
• Recall Tiers (T0–T4). Progressive context loader — persona, horizon, orbit, deep, atlas.
• Watchers. Auto-capture from Claude Code, Codex CLI, Gemini CLI, Copilot CLI, Aider, Antigravity, ChatGPT/Claude web imports, Cursor, Windsurf.
• Chronograph. Bitemporal knowledge graph — entities, relationships, validity windows.
• 27 MCP tools. Full surface exposed via the MCP protocol for any compatible client.

Quick start

# One-shot setup (detects installed AI tools, configures models, installs hooks)
memgentic init

# Run the MCP server + the file watcher in one process
memgentic serve --watch

# Semantic search from the CLI
memgentic search "react performance tips"

# Manual store
memgentic remember "Fixed JWT expiry bug: iat claim was missing"

See the project README on GitHub for architecture, full feature matrix, MCP tool reference, and benchmarks.

License

Apache 2.0. See LICENSE.