Experience memory for production AI agents
Retrack builds experience for AI agents.
Your agent starts every task from scratch. Retrack checks which runs worked, turns them into playbooks and warnings, and recalls them before the next similar task, so the same mistake doesn't happen twice.
A Claude Haiku 5.5 support agent (LangGraph, 11 tools) on 16 paired requests, with and without memory; the open test lives in the connector's examples. Full results and limits
How it works
Every run makes the next one better
- 01
Capture
One line records every run: the task, each tool call and result, the answer, tokens and model.
- 02
Verify
Tests, events and user feedback decide whether a run worked. Without them, an AI judge scores it.
- 03
Distil
Verified successes become playbooks. Failures become warnings, with the fix that worked.
- 04
Recall
Before the next similar task your agent gets the right playbook, only when it clearly matches.
Protect runs through all four: a memory firewall keeps poisoned or untrusted runs from ever becoming memory.
Five-minute setup
Three lines in your LangGraph agent
Install the open-source connector, switch it on where your agent starts, and recall memory before each task. Nothing else about how you call your agent changes.
Capture never raises into your code and recall fails open after 1.5 seconds, so Retrack can never take your agent down.
Follow the quickstartpip install retrack-ai
export RETRACK_API_KEY="rtk_live_…"
One memory, four jobs
Learn, protect, test and migrate from the same store
Playbooks
The steps that worked for a kind of task, with conditional steps and pitfalls learned from real runs.
Failure signatures
Errors your agent has hit before, with the fix that worked. A circuit breaker stops repeats it can't recover from.
Memory firewall
Runs fed by untrusted input never become memory unchecked. Prompt injections are quarantined, not learned.
Regression tests
Every playbook and known failure becomes a test. Run retrack test in CI before you ship a prompt change.
Model migration
Replay memory on a new model before a forced upgrade and see, per playbook, what still works.
Measured, not estimated
A small holdout gets no memory, so the dashboard shows the difference memory actually makes.
Not another user-memory API
Remember what worked, not just what's true
User and document memory
Stores facts about users and content: preferences, past conversations, documents. Useful for chat assistants that need to know who they are talking to.
Experience memory (Retrack)
Stores what worked and what failed for your agent: verified procedures and known failures with fixes, recalled before similar tasks, tested in CI and checked before model upgrades.
Questions
Frequently asked questions
What is Retrack?
+
Retrack is experience memory for production AI agents. It captures every agent run, verifies whether it worked, turns successes into playbooks and failures into warnings, and recalls them before the agent's next similar task. A memory firewall keeps untrusted runs out of memory.
What is agent memory, and which kind does Retrack provide?
+
Agent memory is what an AI agent keeps beyond a single model call: the conversation, facts about users, past runs and learned procedures. Retrack provides experience memory, the procedures that worked and the failures to avoid for an agent's own tasks, which makes task agents more reliable over time.
How is Retrack different from Mem0, Zep or Supermemory?
+
Those products remember facts about users and documents. Retrack remembers what worked and what failed for your agent: verified procedures (playbooks) and known failures with their fixes. It also turns that memory into regression tests and a model-migration check, and measures its own effect with a holdout group.
Which agent frameworks does Retrack support?
+
LangGraph and LangChain agents in Python 3.10 or newer, through the open-source retrack-ai connector (MIT licence). One line, retrack.init(agent=...), records every run.
Will Retrack slow my agent down?
+
No. Capture runs in the background and never raises into your code. Recall gives up after 1.5 seconds by default and returns empty memory, so your agent always runs.
How does Retrack stop memory poisoning?
+
Runs that touched untrusted input are tainted and never become active memory without passing the firewall's checks or your review. New playbooks start as candidates you approve, and only runs from trusted tools with real success signals can promote memory on their own.
What happens to secrets and personal data in my agent's runs?
+
The connector strips obvious secrets (API keys, tokens, passwords) on your machine before anything is sent, and the server redacts again before storing. Each project's data is isolated by row-level security in the database.
Give your agent experience.
Sign in with GitHub, copy your key, and the next run starts teaching.