Operator Memory gives a coding agent a durable workspace of Markdown documents that it writes as it works: specs, decisions, standards, research, and lessons. Each new session starts from what the previous one learned instead of re-exploring the codebase and re-learning the architecture.
Every session runs the same loop. The agent consults its knowledge (instructions, a codebase index, specs, and guides), does normal development work, then updates the documents to record what changed. When project facts change, it edits the canonical document rather than appending a new memory record, and it needs no background agents, embeddings, vector database, or model configuration.
Knowledge lives in three partitions: private project knowledge that stays on your machine, shared project knowledge published with the repository and tracked in Git, and personal rules used across projects. Setup happens in conversation with the agent, and on an existing project the knowledge starts sparse until the agent writes specs for the modules you touch.
Supported harnesses are Claude Code, Codex, OpenCode, Pi, DeepSeek Harness, and Kiro.
Features
- Automatic documentation: the agent records specs, decisions, research, and lessons as part of ordinary work
- Readable memory: memory is Markdown documents you can read, edit, and delete
- Consult, build, update loop: each session starts from existing knowledge and ends by updating it
- Partitions: private project knowledge, shared project knowledge, and personal cross-project rules
- Team sharing: shared documents travel with the repository through Git
- No infrastructure: no background agents, embeddings, vector database, or model configuration
- Harness adapters: Claude Code, Codex, OpenCode, Pi, DeepSeek Harness, and Kiro
