Retrack FAQ
Answers about Retrack: supported frameworks, latency, outages, first playbooks, missing memory, stored data, memory poisoning, open source and pricing.
Updated
Which frameworks and languages are supported?
LangGraph and LangChain agents in Python 3.10 or newer, through the retrack-ai connector. Agents in other languages or frameworks can use the HTTP API directly.
Does Retrack slow my agent down?
No. Capture happens in the background after a run and never raises into your code. Recall has a timeout (1.5 seconds by default) and returns empty memory if it is reached, so your agent never waits longer than you allow.
What if Retrack is down?
Your agent keeps running. Recall returns empty memory, and the connector retries sending each run a few times with backoff in the background. If Retrack stays unreachable, those runs are dropped rather than slowing your agent or growing its memory use.
When will my agent get its first playbook?
After a handful of successful, verified runs of the same kind of task. New playbooks start as candidates: approve them in the dashboard under Memory, or let them switch on by themselves after 3 trusted runs with a hard success signal. See Playbooks.
Why is my agent not getting memory?
Check, in order:
RETRACK_API_KEYis set in the environment where the agent runs.- You call
retrack.recall(task)and putmemory.as_prompt()in the prompt. - The agent has an active playbook for that kind of task. Candidates are not recalled.
- The run is not in the holdout group (
memory.holdoutisFalse). - The task is similar enough to the playbook's: the dashboard's Recalls page shows each recall and its scores.
What data does Retrack store?
Each run's task, steps (tool names, arguments and results), final answer, tokens, cost, model and tool versions, after redaction. Secrets are stripped by the connector before sending and again by the server; personal data such as emails, phone numbers and card numbers is redacted before storage. Each project's data is isolated by row-level security. Details are in the privacy policy.
Can someone poison my agent's memory?
Retrack is designed so they cannot do it silently. Runs that touched untrusted input never become active memory without passing the firewall's checks or your review, prompt injections are quarantined, and every playbook is traceable to its source runs. See the memory firewall.
Is the connector open source?
Yes. retrack-ai is MIT licensed. The Retrack service it connects to is a hosted product.
Does Retrack use my data to train models?
No. Your data is not used to train AI models, by Retrack or by the AI providers it uses. An AI model judges runs without a signal, writes playbooks and checks memory for injections, and your runs are used only for your own project's memory.
How much does it cost?
Retrack is in early access. Sign in with GitHub to start. Each project gets a daily AI budget for judging runs and writing playbooks; above it, runs are still recorded.