Give your agents
context that lasts.

Venom keeps project knowledge in Brain and reusable instructions in Skills. Your agents can retrieve both through MCP, so each new session can continue with the context already in place.

Get early access How it works

Shared memory and skills for teams working with agents.

brains/alex/tasks/attention-window/todo.md

ALEX / ATTENTION WINDOW

Does full attention earn its compute?

The default SSSL pattern alternates three short attention windows with one full window. Compare it with L: full attention at every layer.

Working notes

✓ Located WINDOW_PATTERN and the layer-window calculation in train.py.
✓ Confirmed the final layer always uses full attention.
□ Run the unchanged SSSL baseline, then the L candidate.
□ Compare validation bits per byte and tokens processed.

Next step

Change WINDOW_PATTERN only. More context may help prediction but reduce training throughput.

Based on autoresearch / train.py ↗

Project context → task → experimentAdapted examples

Keep the investigation, not just the answer.

Store the code paths, rejected approaches, and open questions behind a task. The next agent can see how you got here.

Make team expertise reusable.

Publish the review methods and experiment protocols your team trusts. Reuse them across projects, people, and agents.

Use it from your agent.

Connect through MCP to read project notes, retrieve published skills, and update task knowledge as the work progresses.

How it works

From project context to the next experiment.

An agent opens the task notebook, retrieves the experiment method, and saves the outcome alongside the work. Follow the example below, based on Karpathy’s autoresearch project.

01 / Read

Read the experiment notebook.

The agent finds the attention-window hypothesis, code map, and fixed compute budget in Brain.

brains/alex/tasks/attention-window/todo.md

02 / Work

Apply the experiment method.

It retrieves the published skill: establish a baseline, hold evaluation fixed, and compare the candidate.

Published skill / run-an-experiment

03 / Write back

Save the result with the task.

It records the run’s code revision, validation score, and outcome. The next session can inspect the evidence.

brains/alex/tasks/attention-window/experiments/results.md

Meet the MCP and plugin

Skills / Reusable expertise

The methods you trust. Available to your agents.

A code review that identifies unnecessary abstractions. An experiment protocol with a fixed baseline. An evaluation method that keeps comparisons fair. Publish the instructions once and make them available to your team’s agents.

Import an existing SKILL.md or build a skill with supporting files. Review it as a draft, then publish it through your organization’s MCP connection. Explore the adapted Ponytail and autoresearch examples here.

Explore Brain and Skills

skills/ponytail/SKILL.md

SHARED WITH YOUR TEAM

Ponytail

Published · Available through MCP

Choose the smallest implementation that solves the actual problem. Start by checking what the codebase and platform already provide.

Method

Reuse an existing helper before creating another abstraction.
Prefer the standard library or a native platform feature over a new dependency.
For a bug, trace the callers and fix the shared cause.

Skill source

skills/ponytail/SKILL.md

Adapted from Ponytail ↗

Reusable instructions, shared through MCPAdapted examples

Give your agentscontext that lasts.