AtMem
A local-first, auditable memory control plane for AI agents that records exact host-observed text and verifiable execution evidence, authorizing and scoping what gets stored, injected, corrected or deleted, with integrations for OpenClaw, MCP, Pydantic AI and LangChain/LangGraph.
AtMem is a host-neutral memory control plane for AI agents: rather than an agent framework itself, it is infrastructure that other frameworks install to get governed, long-term memory. It records what a connected agent actually observed and supplied, keeping every claim limited to that source material, and is the sole component authorized to store, scope, inject, correct or delete a memory - a companion component called AtBot proposes and ranks candidate memories but never writes them directly.
Installation is a single Python package that also pins a matching AtBot release and sets up an always-present local vector index; during setup, a developer chooses between a local Ollama model, a local OpenAI-compatible model, a hosted provider or a deterministic fallback for memory intelligence, with API keys kept only in environment variables. Connecting to OpenClaw is fully managed by the AtMem CLI, which installs the matching bridge, discovers agents and workspaces, and starts in a safe shadow mode before verifying the connection; other frameworks connect through MCP, Pydantic AI or LangChain/LangGraph integrations.
A separate component, AtFlows, adds workflow attempt visibility and grouped event charts on top of the same memory layer.