Recall and the confidence gate
How Retrack finds the right playbook for a task: hybrid search, a confidence gate that blocks wrong memory, warnings, plan replay and fail-open timeouts.
Updated
Wrong memory is worse than no memory. A playbook from a different kind of task can push an agent down the wrong path, so Retrack only injects a playbook when it clearly applies.
Calling recall
memory = retrack.recall(task) # before the task
prompt = memory.as_prompt() # "" when there is nothing to addOptions: max_playbooks (default 3, up to 10), timeout (seconds, default 1.5), model_id and tool_versions to override what you set in retrack.init.
Recall fails open: on a timeout, a network error or a missing key it returns empty memory and logs a warning. It never raises, so it can never stop your agent.
How candidates are found
Retrack searches your agent's active playbooks two ways at once, by meaning (embeddings of the task) and by keywords, and merges the results. If embeddings are slow or unavailable, recall falls back to keywords alone rather than waiting. Each playbook is matched against its own task and the other wordings it was verified on.
The confidence gate
Every candidate gets a score:
score = similarity × confidence × success_rate × freshness × version_match| Factor | What it measures |
|---|---|
| similarity | How close the task is to the playbook's kind of task |
| confidence | How strong the evidence was for the runs that taught it |
| success_rate | How often runs that used it succeeded |
| freshness | Decays with time since its last success (30-day half-life) |
| version_match | 1.0 when validated on your current model and tool versions; lower when only tools match; 0 after a breaking tool version change |
Only playbooks above your agent's threshold (0.72 by default) are returned, near-copies are dropped, and a playbook whose similarity is below the embedding model's floor is never returned, whatever its other factors.
What you get back
| Field | Meaning |
|---|---|
memory.playbooks | Gated playbooks, best first |
memory.warnings | Up to 5 relevant failure signatures |
memory.holdout | True when this run is in the holdout group and gets no memory |
memory.replay_plan | Set only when plan replay is allowed (below) |
Linking recall to the run
retrack.recall(task) links itself to the next run automatically, so Retrack can count whether the playbook helped. With the explicit client, pass the same run_id to rt.recall(...) and to the graph's config (see the Python SDK).
Plan replay
For very well-proven playbooks, recall can offer a plan to follow directly, skipping planning. It is offered only when the task is a near-exact match, the playbook has at least 20 uses with at least 90% success, and it was validated on your exact model and tool versions. Playbooks with conditional steps are never replayed. A run that follows the plan reports it with rt.callback(replay_of=playbook_id); if it fails, the playbook goes back to candidate for review.
Speed
Measured from Vercel's Mumbai region, recalls that search by meaning took about 370 to 530 ms of server time; keyword-only recalls take a few tens of milliseconds. Your agent never waits longer than the timeout you set.