aerovato/operator-memory

Operator Memory is a cross-platform, self-improving context engine for coding agents that bundles memory, documentation, codebase indexing, and reusable skills into one integrated system. Rather than relying on embeddings, vector databases, background curators, or model configuration, it persists specs, decisions, research notes, and lessons as plain Markdown, organized across three durable partitions: .operator/ for private project knowledge, .operator-shared/ for documents published with the repository, and ~/.operator/user/ for personal rules reused across projects. Each session runs a tight consult–build–update loop so that what an agent learns today becomes a file the next session can open directly, instead of being reconstructed from lossy RAG chunks.

The repository is packaged and installed through a dedicated command-line tool, @aerovato/operator-helper, which can be installed globally via npm install --global @aerovato/operator-helper or bun add --global --minimum-release-age 0 @aerovato/operator-helper@latest. After the Helper is in place, users run an adapter command such as operator-helper install opencode-v2, operator-helper install opencode, operator-helper install pi, operator-helper install codex, or operator-helper install code-puppy to mount Operator onto their chosen harness, while Claude Code support remains on the public roadmap pending research. First-time setup is a conversation: users invoke /operator:user-init once for the global user partition and /operator:project-init inside each new project so the agent can scaffold Operator, migrate existing documents, and index the repository before normal work begins.

Operator Memory targets development teams and individual engineers who want coding agents to retain architectural understanding, decisions, and conventions across sessions, and to share that knowledge through Git-tracked documents. Dependencies are intentionally minimal: only Node.js or Bun is required, with no embeddings, no vector store, no background dreamers, and no model setup, and the project is released under the BSD-3-Clause license. Practical limitations include the fact that some agents hesitate to create, consolidate, or split documents, so users must occasionally steer them with prompts like "write a spec for this feature" or "consolidate these overlapping documents," and additional harnesses such as Claude Code are not yet available at the time of writing.