Changelog
What is new in Retrack: the connector, the dashboard, playbook learning, failure warnings, the memory firewall and the public documentation.
Updated
5 November 2026
- Public website and documentation: concepts, guides and reference for the connector, the CLI and the HTTP API, with Markdown copies of every page for AI assistants.
- One-line setup (connector 0.2.0):
retrack.init(agent=...)records every LangGraph and LangChain agent run with no callback to pass, andretrack.recall(task)links itself to the next run. The dashboard's Copy setup prompt gives an AI coding assistant the same steps. - General-purpose tools: agents built around a shell, SQL or HTTP tool get memory, warnings and fixes per command.
- Honest outcomes for crashes: a crash in a tool or the agent's code counts as a failure; a model-provider outage does not.
4 November 2026
- Playbooks learn from every verified run. New tools a task needed become conditional steps, recovered errors become pitfalls, and a person approves changes to playbooks in use.
- One playbook per kind of task, recalled however the task is worded.
- Warnings explain the failure and the fix: the tool's own message and which inputs made the retry work.
- Maple & Main results: avoidable mistakes −80% on a real support agent. See benchmarks.
3 November 2026
- Sign in with GitHub.
- First-run guidance in onboarding: from the first run to approving the first playbook.
- Connector 0.1.1: the API moves to
api.retrack-ai.online.
2 November 2026
retrack-ai0.1.0 on PyPI (MIT licence).
30 October 2026
- Approve button: project owners and admins can activate a candidate playbook from its page.
- Warnings by tool and kind of task: failure warnings reach agents that version their APIs rather than their tools.
- Ops Desk benchmark: a harder world with unwritten API rules.
21 October 2026
- Dashboard: overview with memory lift against the holdout, runs explorer, memory board with version diff, review queue, recall inspector, regression and model-migration reports.
- Fleet sharing: propose a playbook to another agent; it goes live only after approval, and any version can be rolled back.