The maintainer is unusually direct about it: the current tool surface is being cut in full, with no data import into what replaces it. That is worth knowing before you build a workflow on `memory_record`. Run it if the cross-session recall earns its keep this quarter, and keep anything you would miss somewhere the rewrite cannot reach.
A Go memory service for agents, running as a container with its own vector store. The shipped build records strategies that worked, error patterns and their fixes, and session checkpoints, and searches them back at the start of the next task. The README announces a complete rewrite aimed at knowledge shared between engineering teams and their agents — Incidents, Decisions and Standards attached to an Org, Team or Project — with no migration and no compatibility shim from what runs today.
- Memory: `memory_search`, `memory_record`, `memory_feedback`, `memory_outcome`, `memory_consolidate` and `memory_consolidate_session` — memories carry an outcome of success or failure and a confidence score that moves with feedback
- Checkpoints: `checkpoint_save`, `checkpoint_list` and `checkpoint_resume` for picking a long task back up
- Error patterns: `remediation_search`, `remediation_record` and `remediation_feedback`
- Sub-task isolation with token budgets: `branch_create`, `branch_return`, `branch_status`
- Code and conversation indexing: `repository_index`, `repository_search`, `semantic_search`, `conversation_index`, `conversation_search`
- Diagnostics: `troubleshoot_diagnose`, `reflect_report`, `reflect_analyze`
The container ghcr.io/fyrsmithlabs/contextd:latest, run over stdio with a named volume mounted at /data. Embeddings run locally through fastembed, so no cloud key is needed; the vector store is embedded by default and Qdrant is optional, configured with QDRANT_HOST and QDRANT_PORT. CONTEXTD_DATA_PATH defaults to /data.
One command — contextd --mcp --no-http