Zep Agent Skills · Free until Jan 4, 2027

Build agents that
learn from experience

Zep Agent Skills is the learning layer for enterprise agents, turning agent tasks into skills that improve completion rate and reduce cost.

Start building

Why Zep Agent Skills

More reliable, lower-cost agents

For complex tasks, Zep Agent Skills learns from each run and improves the skills automatically.

Complete more tasks reliably

Your agent reuses the steps that worked before. In our testing, task success rose by up to 18.8 points.

See the benchmark results

Reduce token use and cost

Skills from a strong model let a lower-cost model do the same task. In our tests, completed tasks used up to 44.6% fewer tokens.

See the token results

Finish faster with less reasoning

With a known procedure, your agent needs less reasoning. Each model with skills at low reasoning beat its maximum-reasoning score without skills.

See the reasoning chart

Ready to start? Install the Zep v4 SDK and send your first Traces.

Go to Get started

The problem

Every run starts from zero

Your agent has to discover the same procedure on each run and makes the same mistakes again. When the run ends, the agent loses what it learned.

Task · Reconcile a failed paymentRepeated in every run

Run

Repeated in every run

Finds the cause again

Result

MONrun 0412

retry charge✕ declined

inspect intent

card expired

open case

resolved14 steps

TUErun 0463

retry charge✕ declined

retry charge✕ duplicate charge

refund duplicate

inspect intent

card expired

resolved21 steps

WEDrun 0518

retry charge✕ declined

retry charge✕ gateway timeout

inspect intent

card expired

open case

resolved17 steps

MON · run 0412resolved · 14 steps

retry charge✕ declined
retry charge✕ declined
inspect intent
card expired
open case

TUE · run 0463resolved · 21 steps

retry charge✕ declined
retry charge✕ duplicate charge
refund duplicate
inspect intent
card expired

WED · run 0518resolved · 17 steps

retry charge✕ declined
retry charge✕ gateway timeout
inspect intent
card expired
open case

The solution

With Zep Agent Skills, agents improve with every success and failure

Zep compiles the steps that worked in your agent's past runs into a skill. The agent follows the skill, and each new run updates the skill.

Past runs

PASSrun 0518
FAILrun 0463
PASSrun 0531
FAILrun 0490

Agent Skills

SKILL · V5

Reconcile a failed payment

1 Check the payment intent first
2 Retry once, with an idempotency key
3 On a second failure, open a case

Next run

check intent
update payment method
retry once
resolved first time

Each run refines the skill

Get started

Integrations

Works with the agent stack you have

Zep Agent Skills works with any agent framework. You can import your existing traces or send Traces with the Zep SDK.

Agent frameworks

LangGraph
OpenAI Codex SDK
Claude Agent SDK
Google ADK
Mastra
Eve

Import traces from Braintrust Import traces from LangSmith Send Traces with the Zep SDK

Trace connections

BraintrustAvailable
LangSmithAvailable
LangfuseComing soon

Benchmark results

More tasks completed on every benchmark we tested

Zep built skills from Claude Opus 5.5 runs. Three lower-cost models then ran the same tasks with and without the skills, on three enterprise benchmarks.

Skills make cheaper models more capable

Tasks completed · EnterpriseOps-Gym · low effort

GPT-6 Luna27.7%
+ skills37.9%
Claude Haiku 5.533.9%
+ skills45.3%
DeepSeek V4.1 Flash48.3%
+ skills56.8%
Claude Opus 5.5Max effort, no skills52.8%

Tasks completed

+18.8 pts

Agents with skills completed more tasks in all 26 configurations that we tested.

GPT-6 Luna on AppWorld, low reasoning effort

Tokens per completed task

−44.6%

Fewer tokens in all 8 τ³-bench banking configurations.

Claude Haiku 5.5 on τ³-bench banking, medium reasoning effort

How we tested: each model ran the same tasks as Claude Opus 5.5, and most of the gain was on tasks that Claude Opus 5.5 solved. Claude Opus 5.5 ran each task once. On AppWorld, token use went up in 7 of 9 configurations because our test harness put all 62 skills in every prompt.

Results · October 2026

Read the paper

Start building

How Agent Skills works

Capture, compile, curate, review, serve

Each verified outcome feeds the next compilation. The agent always reads the current admitted version.

Capture

Each task run becomes a Trace: the objective, the ordered events, and the verified outcome.

Compile

Zep proposes a procedure from the verified runs of one task family.

Curate

Zep checks each proposal, then adds it to a new version or rejects it.

Review

Automatic scans and an evaluation run before Zep admits a version. You can add member review.

Serve

The agent reads the current admitted skill through MCP or the Zep SDKs.

Serve

Skills reach the agent at task time

Connect an MCP-compatible agent with no change to its code, or call the Zep SDKs from your application.

Context MCP Server

Bind an Agent to the MCP connection of your project. The agent gets tools to search skills, read a skill, and report each use and its outcome.

agent → Zep Context MCP Server

search_agent_skills("failed invoice payment")

→ payment_recovery · v5

read_agent_skill("payment_recovery")

→ 4 steps, 1 recovery

report_agent_skill_use(outcome="succeeded")

Use skills through MCP

Zep SDKs

Search before the task, put the skill in the prompt, then record the use and its outcome.

skills.py

results = client.agent.skill.search(    agent_uuid=agent_uuid,    query="customer reports a duplicate charge",    task_family="billing.support",    limit=3,)skill = results.items[0]  # put skill.markdown in the prompt use = client.agent.skill.use.create(    agent_uuid=agent_uuid,    skill_uuid=skill.uuid_,    search_id=results.response.search_id,    trace_uuid=trace.uuid_,    usage="used",) client.agent.skill.use.add_outcome(    agent_uuid=agent_uuid,    skill_uuid=skill.uuid_,    use_uuid=use.uuid_,    outcome="succeeded",)

skills.ts

const results = await client.agent.skill.search(agentUuid, {  query: "customer reports a duplicate charge",  taskFamily: "billing.support",  limit: 3,});const skill = results.data[0]; // put skill.markdown in the prompt const use = await client.agent.skill.use.create(agentUuid, skill.uuid!, {  searchId: results.response.searchId!,  traceUuid: trace.uuid!,  usage: "used",}); await client.agent.skill.use.addOutcome(agentUuid, skill.uuid!, use.uuid!, {  outcome: "succeeded",});

Skills and fine-tuning

Task knowledge without a training pipeline

Fine-tuning is costly, slow, and brittle, and each process change needs a new training run for one model. Skills need no training: on ALFWorld, skills removed 57% of Qwen3-8B failures, and RSD fine-tuning removed only 37%.

Operating effort: Fine-tuning
Curate datasets, run training jobs, maintain evals, and deploy and host each model version.
Agent Skills
Skills are compiled from runs the agent already makes. No datasets, training jobs, or model hosting.
Keeping current: Fine-tuning
Updates arrive on a retraining cycle. A new base model means fine-tuning again.
Agent Skills
Skills update as new runs arrive, and carry over when you change models.
Inspection and rollback: Fine-tuning
Learned behavior is in the weights. Reverting means redeploying an earlier model.
Agent Skills
Each skill is a readable procedure with cited steps. Reverting means serving an earlier version.

Topic	Fine-tuning	Agent Skills
Operating effort	Curate datasets, run training jobs, maintain evals, and deploy and host each model version.	Skills are compiled from runs the agent already makes. No datasets, training jobs, or model hosting.
Keeping current	Updates arrive on a retraining cycle. A new base model means fine-tuning again.	Skills update as new runs arrive, and carry over when you change models.
Inspection and rollback	Learned behavior is in the weights. Reverting means redeploying an earlier model.	Each skill is a readable procedure with cited steps. Reverting means serving an earlier version.

Governance

Every skill is readable, cited, and versioned

You can see what the agent will do before it does it, where each step came from, and how the procedure changed.

Boundaries

Skills stay in the security domain of their Agent. A member can share a skill version with another Agent in that domain, and that Agent admits the copy before use.

Provenance

Every step of a skill links to the runs that it came from.

Versions

Versions are immutable. Compare two versions, or serve an earlier version.

Review

Zep admits a version after automatic checks, or after member review for the task families that you select.

Get started

Install the Zep v4 SDK, send or import a few verified runs, and search for the skill that Zep compiles.

Create a Zep account

Install the Zep v4 SDK

The Agent Skills methods are in the Zep v4 SDKs only. The Zep 3.x SDKs do not have them.

Import or capture runs

Import traces from Braintrust or LangSmith, or send Traces from your agent with the Zep SDK.

Serve skills at task time

Search skills before each task, and record each use and its outcome.

Python

pip install --upgrade zep-cloud

TypeScript

npm install @getzep/zep-cloud

Documentation

Zep Agent Skills

Knowledge and skills, in one context layer

Context graphs give your agent current facts. Agent Skills gives it the procedures that worked. Try Agent Skills free for 60 days.

Start building

Build agents thatlearn from experience

More reliable, lower-cost agents

Complete more tasks reliably

Reduce token use and cost

Finish faster with less reasoning

Every run starts from zero

With Zep Agent Skills, agents improve with every success and failure

Works with the agent stack you have

More tasks completed on every benchmark we tested

Skills make cheaper models more capable

Capture, compile, curate, review, serve

Capture

Compile

Curate

Review

Serve

Skills reach the agent at task time

Context MCP Server

Zep SDKs

Task knowledge without a training pipeline

Every skill is readable, cited, and versioned

Boundaries

Provenance

Versions

Review

Get started

Create a Zep account

Install the Zep v4 SDK

Import or capture runs

Serve skills at task time

Knowledge and skills, in one context layer

Build agents that
learn from experience