Your agent fails
You restart it
It fails at the exact same thing again
Sound familiar
The problem every AI team hits
Every team building autonomous agents eventually rebuilds the same three things
Memory so the agent remembers what failed last time
Retry logic so it does not loop forever on the same broken approach
Orchestration so multiple agents do not step on each other
You build it It works You start the next project and build it again from scratch
There is no standard layer for this Until now
Introducing NEXUS
One line install Works with any agent Gets smarter over time
pip install cognicore env
import cognicore as cc
env = cc.make SafetyClassification Easy v1
agent = cc.AutoLearner
cc.train agent=agent env=env episodes=30
score = cc.evaluate agent=agent env=env episodes=5
What makes it different
Memory that compounds
The more tasks NEXUS handles the better it gets
text
Week 1 0.05 per fix
Week 4 0.02 per fix
Week 8 0.01 per fix
An agent with 6 months of memory on your codebase is fundamentally different from one starting cold
Agent Immune System
Protect any agent from prompt injection jailbreaks and token bombs
python
from cognicore.immune import NexusShield
safe_agent = NexusShield agent=your_agent
Replay and Time Travel
Every decision event sourced Rewind any task to any step Branch and try a different strategy
cognicore replay task abc123
cognicore branch task abc123 step 3 policy minimal
6 Enterprise Integrations
Label a GitHub issue nexus NEXUS fixes it opens a PR automatically
bash
cognicore integrations setup
Live Dashboard
bash
cognicore ui
The research finding that surprised everyone
I ran ablation studies comparing multi agent configurations
Expected more specialized agents equals better results
Actual
minimal Coder Tester only 19 20 solved 0.014
full pipeline 5 agents 18 20 solved 0.009
review first ordering 18 20 solved 0.009
The Reviewer agent costs minus 1 solve rate and plus 9642 tokens
More agents Worse performance More expensive
An offline RL agent trained on 220 trajectories independently confirmed minimal policy wins 89 percent of task states
For developers building AI agents
Stop rebuilding memory from scratch on every project
from cognicore import Memory ReflectionEngine
mem = Memory
ref = ReflectionEngine memory=mem
action reason confidence = ref.suggest_override
null handling
guard fix
For ML researchers
38 built in environments across 6 domains
4 RL agent types with clean interfaces
Ablation infrastructure with statistical rigor
460 plus trajectories exportable for offline RL
SWE bench style evaluation built in
CognitiveMemory with working episodic semantic and procedural layers
from cognicore import Experiment
exp = Experiment
name=memory ablation
env id=SafetyClassification v1
exp.add_variant no memory cc.AutoLearner
exp.add_variant with memory cc.AutoLearner
results = exp.run episodes=50
For CTOs and engineering leads
Self hostable
Open source core Apache 2.0
Token cost tracking built in
Budget controls
Full audit log
GitHub Slack Linear integrations
text
Devin 500 month
NEXUS 3 to 15 month
Numbers
1700 plus downloads in first week
95 percent solve rate on SWE style benchmark
472 tests passing
62 built in environments
153 public API exports
Zero required dependencies for core
6 enterprise integrations
460 plus trajectories stored for offline RL
Try it in 2 minutes
bash
pip install cognicore env
cognicore ui
cognicore integrations setup
python
import cognicore as cc
env = cc.make GridWorld v1
agent = cc.AutoLearner
cc.train agent=agent env=env episodes=50
print
cc.evaluate agent=agent env=env episodes=5
GitHub
github com Kaushalt2004 cognicore my openenv
PyPI
pypi org project cognicore env
Docs
cognicore readthedocs io
Open source Apache 2.0 Solo built Actively maintained
Star the repo if this solves a problem you have hit before
