Agent Experience Graph GitHub

Open source · MIT-0

Agents should not start from zero every time.

Agent Experience Graph turns sanitized execution traces into reusable skills, tools, workflow patterns, and failure warnings—so the next agent starts with evidence, not an empty slate.

For teams taking AI agents from pilots to production. Reuse proven execution patterns · reduce retries · preserve privacy
See the 30-second demo View source
  • 01 Runtime-neutral
  • 02 Privacy-first
  • 03 Python 3 · offline
EXPERIENCE LOOKUP trace://local
NEW TASK

Build markdown API docs ingestion

parse sections extract metadata
RECOMMENDED structured-ingestion evaluation-harness success trace

The vision

Foundation models are shared.
Execution experience isn’t.

Today, agents repeatedly rediscover the same decomposition, tool choices, guardrails, and dead ends. AEG makes those lessons retrievable across tasks and runtimes.

Think of it as workplace memory for agents: “Who has solved something like this before, and what worked?”

Core capabilities

Experience becomes infrastructure.

Small, inspectable building blocks that fit the agent stack you already use.

01

Subtask-aware retrieval

Match what the agent is trying to do—not just the title of the task—against successful and failed prior attempts.

parse markdownmatched evaluate schemamatched
02

Evidence-backed guidance

Recommend skills, tools, lessons, and warnings with outcome context. The agent decides what fits its current constraints.

skilltoolwarning
03

Sanitized by design

Store compact lessons and execution structure—not secrets, customer content, proprietary code, or raw conversations.

✓ task + outcome ✓ tools + lessons × raw private data

Architecture overview

A compact learning loop.

Portable markdown instructions, a neutral manifest, and a small local recommender.

01 INGEST

Sanitized traces

Task, subtasks, outcomes, skills, tools, constraints, and reusable lessons.

02 RETRIEVE

Similarity + evidence

Rank related attempts using task and subtask overlap, with outcome context.

03 RECOMMEND

Skills + warnings

Surface proven tools, workflow patterns, local lessons, and failed paths.

04 LEARN

Next trace

After execution, record a redacted summary so future agents improve.

Works with CodexClaude CodeOpenClawGeneric agents network: false

Example

Watch a task inherit experience.

This interaction mirrors the repository’s bundled trace recommender.

01 CURRENT TASK

Build markdown API docs ingestion

  • Parse markdown into sections
  • Evaluate endpoint metadata extraction

Ready to search sanitized traces

No network required

Quick start

One repository.
No service required.

Clone the project and run the recommender against the bundled trace library. Python 3.9+ is the only runtime requirement.

Installation options for every runtime
terminal 
# Clone the portable capability
git clone https://github.com/yao23/agent-experience-graph.git
cd agent-experience-graph

# Retrieve experience for a new task
python3 scripts/recommend_traces.py \
  --traces assets/example_traces.json \
  --query '{"task":"Build markdown API docs ingestion","subtasks":[{"description":"parse markdown into sections"}]}'

Contribute

Help agents learn
from more than success.

The most useful graph includes what worked, what failed, and why. Contributions should stay portable, testable, and privacy-safe.

Explore the repository
01

Share sanitized traces

Add compact examples with clear outcomes and lessons—never raw private work.

02

Improve retrieval

Experiment with ranking, evaluation, negative evidence, and trace quality signals.

03

Build runtime adapters

Make the same capability easy to load from more agent environments.

04

Design benchmarks

Measure whether experience lowers retries, latency, token cost, and failure rates.