# Retrack > Retrack builds experience for AI agents. It captures every agent run, verifies whether it worked, turns successes into playbooks and failures into warnings, and recalls them before the next similar task. A memory firewall keeps poisoned or untrusted runs out of memory, and the same memory doubles as a regression suite and a model-migration check. Python connector: `pip install retrack-ai` (LangGraph and LangChain). ## Guides - [Agent memory: a complete guide for AI agents](https://retrack-ai.online/guides/agent-memory.md): What agent memory is, the kinds AI agents need, how to add long-term memory to a LangGraph agent, the tools compared, and how to keep memory safe. - [How to improve AI agents in production: 7 steps](https://retrack-ai.online/guides/improve-ai-agents.md): A practical method to make AI agents more reliable over time: verify every run, learn from it, stop repeat mistakes, test every change, measure. - [How to make AI agents cost-efficient long term](https://retrack-ai.online/guides/ai-agent-costs.md): Where AI agent spend really goes and how to cut it over time: failed runs, retry loops, long tool paths and safe moves to cheaper models. ## Docs - [Retrack documentation](https://retrack-ai.online/docs.md): How to give a LangGraph or LangChain agent experience memory with Retrack: install, capture runs, recall playbooks, send feedback and test in CI. - [Quickstart: add memory to a LangGraph agent](https://retrack-ai.online/docs/quickstart.md): Install the retrack-ai connector, switch it on with one line, recall playbooks before each task and report outcomes. Five minutes, Python 3.10+. - [How Retrack works](https://retrack-ai.online/docs/how-it-works.md): The pipeline behind experience memory: capture every agent run, verify it, distil playbooks and failure signatures, recall them, and protect memory. - [Playbooks: procedures your agent learned](https://retrack-ai.online/docs/playbooks.md): What a Retrack playbook contains, how it is learned from verified agent runs, how it gains conditional steps and pitfalls, and when it is recalled. - [Failure signatures and warnings](https://retrack-ai.online/docs/failure-signatures.md): How Retrack fingerprints repeated agent failures, learns the fix that worked, and warns your agent before it makes the same mistake again. - [Recall and the confidence gate](https://retrack-ai.online/docs/recall.md): 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. - [Verifying runs: outcomes and feedback](https://retrack-ai.online/docs/outcomes.md): How Retrack decides whether an agent run worked: tests, business events and user feedback, an AI judge as fallback, and why hard signals matter. - [Memory firewall: stopping memory poisoning](https://retrack-ai.online/docs/memory-firewall.md): How Retrack keeps prompt injection and untrusted input out of agent memory: trust from settings, quarantine, review, provenance and an attack suite. - [Measuring impact with a holdout group](https://retrack-ai.online/docs/holdout.md): Retrack holds out a slice of runs with no memory and compares success rate, tokens and cost, so the effect of memory is measured, not estimated. - [Regression tests for AI agents in CI](https://retrack-ai.online/docs/regression-tests.md): Turn your agent's playbooks and known failures into regression tests and run them with retrack test in GitHub Actions or any CI before you ship. - [Model migration: test a new model first](https://retrack-ai.online/docs/model-migration.md): Replay your agent's memory on a new LLM before a forced model upgrade with retrack migrate, and see per playbook what passes, breaks or needs relearning. - [Known fixes and the circuit breaker](https://retrack-ai.online/docs/protect-tools.md): Wrap LangChain tools with memory.protect so known failures return the fix that worked and repeated failures stop instead of looping. - [Python SDK reference (retrack-ai)](https://retrack-ai.online/docs/python-sdk.md): Reference for the retrack-ai Python connector: init, recall, feedback, flush, the Memory object, the explicit Retrack client and its exceptions. - [CLI reference: retrack test and migrate](https://retrack-ai.online/docs/cli.md): Reference for the retrack command: run memory-derived regression tests and model-migration replays, with options, entry points and exit codes. - [HTTP API reference (v1)](https://retrack-ai.online/docs/api.md): Retrack's HTTP API: send runs, recall memory, report outcomes, manage playbooks, run regression and migration reports, and read holdout metrics. - [Retrack FAQ](https://retrack-ai.online/docs/faq.md): Answers about Retrack: supported frameworks, latency, outages, first playbooks, missing memory, stored data, memory poisoning, open source and pricing. ## Features - [Playbooks: procedural memory for AI agents](https://retrack-ai.online/features/playbooks.md): Retrack turns verified agent runs into playbooks, procedures with conditional steps and pitfalls, recalled only when they clearly apply. - [Failure warnings and known fixes for agents](https://retrack-ai.online/features/failure-signatures.md): Stop AI agents repeating the same mistakes: Retrack remembers each failure and the fix that worked, warns in advance and breaks retry loops. - [Memory firewall: protection from memory poisoning](https://retrack-ai.online/features/memory-firewall.md): Keep prompt injection and untrusted input out of agent memory with taint tracking, injection detection, quarantine, human review and provenance. - [Regression testing for AI agents, generated](https://retrack-ai.online/features/regression-testing.md): Every playbook and recurring failure becomes a regression test. Run retrack test in CI so prompt and code changes cannot undo what agents learned. - [Model migration testing for LLM agents](https://retrack-ai.online/features/model-migration.md): Replay your agent's memory on a new model before a forced upgrade and get a per-playbook report of what passes, breaks or needs relearning. - [Measured impact: a holdout group for agent memory](https://retrack-ai.online/features/holdout.md): A built-in holdout gives a slice of runs no memory, so Retrack shows the real lift in success rate, tokens and cost on your own traffic. ## Use cases - [Memory for AI customer-support agents](https://retrack-ai.online/use-cases/support-agents.md): Retrack teaches support agents your store's rules from real tickets, warns before known mistakes and keeps customer text from poisoning memory. - [Memory for AI operations and back-office agents](https://retrack-ai.online/use-cases/ops-agents.md): For agents calling internal APIs: Retrack turns tool errors into procedures, stops retry loops and turns lessons into CI tests. Tool errors −67%. ## Compare - [Retrack vs Zep: agent skills and experience memory](https://retrack-ai.online/compare/zep.md): Zep is an enterprise context layer that compiles skills from runs; Retrack adds a memory firewall, failure warnings, CI tests and a holdout. - [Retrack vs Mem0: experience memory or user memory](https://retrack-ai.online/compare/mem0.md): Mem0 remembers facts about users from conversations; Retrack learns what worked and failed for an agent's tasks. How they differ, and using both. - [Retrack vs Letta: memory service or agent platform](https://retrack-ai.online/compare/letta.md): Letta is a platform for stateful agents that edit their own memory; Retrack adds verified experience memory to the LangGraph agent you already have. - [Retrack vs LangMem for LangGraph agents](https://retrack-ai.online/compare/langmem.md): LangMem is an open-source memory library with prompt optimisation; Retrack is a hosted service with playbooks, a firewall, CI tests and a holdout. - [Retrack vs Supermemory: task memory or user context](https://retrack-ai.online/compare/supermemory.md): Supermemory builds user and document context for apps; Retrack learns which procedures work for an agent's tasks and keeps that memory safe. ## Glossary - [What is experience memory for AI agents?](https://retrack-ai.online/glossary/experience-memory.md): Experience memory is an agent's long-term memory of what worked and what failed on its own tasks, recalled before similar tasks so it improves with use. - [What is procedural memory in AI agents?](https://retrack-ai.online/glossary/procedural-memory.md): Procedural memory is an agent's stored knowledge of how to do a task, as opposed to facts or past events. How it is learned from verified runs. - [What is an agent playbook?](https://retrack-ai.online/glossary/playbook.md): An agent playbook is a reusable procedure for one kind of task, with steps, conditions and pitfalls, given to the agent before a similar task. - [What is a failure signature?](https://retrack-ai.online/glossary/failure-signature.md): A failure signature identifies one kind of agent failure by tool, error and input shape, with its frequency and the fix that resolved it. - [What is memory poisoning in AI agents?](https://retrack-ai.online/glossary/memory-poisoning.md): Memory poisoning plants instructions in an AI agent's long-term memory through content it processes, so the agent obeys them in later tasks. - [What is indirect prompt injection?](https://retrack-ai.online/glossary/indirect-prompt-injection.md): Indirect prompt injection hides instructions in content an AI agent processes, such as web pages, emails or tool output. Examples and mitigations. - [What is taint tracking for AI agents?](https://retrack-ai.online/glossary/taint-tracking.md): Taint tracking labels agent inputs as trusted or untrusted and follows that label through a run, so untrusted input cannot change agent memory. - [What is a holdout group?](https://retrack-ai.online/glossary/holdout-group.md): A holdout group is a random share of traffic left without a feature, such as agent memory, so its real effect can be measured against the rest. - [What is plan replay for AI agents?](https://retrack-ai.online/glossary/plan-replay.md): Plan replay lets an agent reuse a proven plan instead of planning from scratch, when a task closely matches one it has solved reliably before. - [What is a circuit breaker for AI agents?](https://retrack-ai.online/glossary/circuit-breaker.md): A circuit breaker stops an AI agent from retrying a tool call that cannot succeed, ending retry loops before they burn tokens and time. - [What is regression testing for AI agents?](https://retrack-ai.online/glossary/agent-regression-testing.md): Regression testing for AI agents replays known tasks after a change and checks outcomes and tool calls, so prompt or model changes cannot break them. - [What is LLM-as-a-judge?](https://retrack-ai.online/glossary/llm-as-a-judge.md): LLM-as-a-judge uses a language model to evaluate an agent's output when no automatic check exists. Its strengths, weaknesses and safe use. - [What is model migration for LLM agents?](https://retrack-ai.online/glossary/model-migration.md): Model migration is switching an AI agent to a new language model while keeping its behaviour on real tasks. Risks and a safe process. ## More - [Benchmarks: agent memory on real Claude agents](https://retrack-ai.online/benchmarks.md): Measured results of Retrack on Claude Haiku 5.5 agents: avoidable mistakes −80% on a support agent, tool errors −67% on an ops agent, and the limits. - [Changelog](https://retrack-ai.online/changelog.md): What is new in Retrack: the connector, the dashboard, playbook learning, failure warnings, the memory firewall and the public documentation. - [About Retrack](https://retrack-ai.online/about.md): Retrack builds experience for AI agents: memory earned from verified runs, recalled only when it applies, protected from poisoning and measured.