memographix

# Memographix Memographix gives AI coding agents local memory for your repo. Install it once, run setup once, and keep using Codex, Claude, Cursor, Copilot, Gemini, OpenCode, Aider, Windsurf, or any MCP client normally. Memographix retrieves fresh context before work and captures useful task memory after work. ## Quick Start Run this inside the repo where you want AI-agent memory: ```bash pipx install memographix mgx setup mgx doctor --live mgx verify-agent --repair mgx savings ``` `pipx` avoids system Python conflicts on macOS and Linux. The PyPI package includes the CLI, local indexer, MCP server, and agent integration support. `mgx setup` creates `.memographix/`, indexes the repo, writes MCP config for supported agents, registers the repo for global MCP routing, and installs project agent rules. For Codex it also installs a small global Codex skill so new chats know to check Memographix before repo work. If you are already inside a virtual environment or CI job, this also works: ```bash python -m pip install memographix ``` PyPI resolves the latest release automatically, so install commands stay versionless. ## Daily Use Use your AI agent normally. In strict mode the agent asks Memographix for fresh context before repo work, captures useful results after work, and marks old memory stale when evidence files change. ```bash mgx doctor --live mgx verify-agent --repair mgx heal mgx guard mgx savings --since 30d ``` `mgx doctor --live` verifies that the MCP server starts, expected tools are available, and the router can resolve this repo. It does not prove your active agent has actually called Memographix yet. `mgx verify-agent` gives you a short prompt to paste into the agent and passes only after that agent performs a real `resolve_task` and `capture_task`. `--repair` lets Memographix refresh stale MCP config before verification. Restart agents after setup if they were already open. This restart matters because Codex and other agents load MCP tools/rules at chat startup. Control it per repo: ```bash mgx status mgx disable --reason "not needed here" mgx enable ``` Disabled repos keep existing memory but automatic agent calls return no context and save nothing. Re-enabling refreshes the index before Memographix is used again. If savings are all zero, Memographix now tells you whether no agent tool calls have been recorded yet. Run `mgx doctor --live`, `mgx verify-agent`, restart the agent, and either open the chat from the repo or mention a registered repo name. `mgx guard` warns when Memographix saw no MCP usage, saw context retrieval without later capture, or sees modified files without any capture event. Other useful commands: ```bash mgx repos mgx doctor --live --repair mgx heal ``` `mgx heal` is safe to run repeatedly. It reruns setup, refreshes MCP command paths, reinstalls project rules, registers the repo, and keeps existing memory. ## Proof On the pinned Kubernetes benchmark, Memographix wins the repeated-task metrics it is designed for: lower repeated tokens, faster repeated recall, stale-evidence safety, and the best deterministic quality score. See [Benchmarks](BENCHMARKS.md) for the public corpus, exact commands, honest losses, and unavailable-tool notes. ## Boundaries Memographix does not upload your code, save full chat transcripts by default, treat stale memory as correct, or install benchmark tools in the runtime package. ## Docs - [Benchmarks](BENCHMARKS.md): public Kubernetes results and claim policy. - [Security](SECURITY.md): local privacy, skipped secrets, and sandbox safety. - [Agent integrations](docs/AGENT_INTEGRATIONS.md): MCP and agent setup. - [Architecture](docs/ARCHITECTURE.md): Python/Rust design and storage model. - [Repeat task memory](docs/REPEAT_TASK_MEMORY.md): capsules and freshness. - [Contributing](docs/CONTRIBUTING.md): local development and test expectations. - [PyPI release](docs/PYPI_RELEASE.md): trusted publishing and release checks.