Stash

This repository provides the open-source foundation: trace collection, persistent knowledge, and reusable skills, accessible through a Python SDK, REST API, MCP, and CLI. It works alongside your existing agents and models.

flowchart LR
    A[Agent runs and human corrections] --> B[Capture sessions via hooks or API]
    B --> C[Curate durable knowledge]
    C --> D[Retrieve knowledge or package skills]
    D --> E[Use in the next agent run]
    E --> A

Stash's broader work focuses on extracting reliable feedback signals from messy production traces and using them to improve prompts, skills, and ultimately model weights. The reward-model product docs cover trace review, reward-model training, and skill optimization with GEPA. The implementation is included in backend/services/rm/ and rm_worker/. Our research focuses on improving the reliability of feedback extracted from production traces.

The reward-model experiment is enabled for accounts created after the rollout. Existing accounts retain their current experience. Access is enforced in both the UI and API; see the rollout guide.

Reward models in action

Continuously improve prompts. The Optimization tab connects a personal reward model to future agent runs through MCP. It proposes instructions, compares current and candidate prompts on real tasks, and tracks reward alongside a chosen business metric. See continuous prompt optimization for setup, promotion criteria, pause/rollback and prototype limits.

Train from feedback in your traces. Stash derives preference pairs from reviewer comments and user reactions, including approval, disappointment, and corrections. A classifier attributes each judgment to a response and keeps the source quote; unclear reactions are excluded from training. You can add comments by highlighting a response. There is no separate Auto mode or rating control in the UI; existing explicit ratings can be supplied through the API. Training needs at least two grounded comparisons. User reactions are used when available; an AI assessor also evaluates response quality without manual annotations or reactions.

The product demo below follows a refund-support agent from reviewer feedback to a reusable skill. These screenshots show an earlier interface with demo data; the current controls are described below.

1. Review a trace. Highlight a response and explain what should change. Feedback already present in the conversation is extracted automatically during training.

2. Train a reward model from traces. Select traces and choose Create new reward model. The worker extracts supported preferences, records their source, and trains the model. AI-generated comparisons receive a separate review for grounding and preference quality. View learning distinguishes user feedback from AI judgments, shows the supporting response or quote, and records whether each finding was included in training. AI judgments are model preferences, not measured customer satisfaction.

3. Turn the reward signal into a skill. Choose View skill on a trained model. It opens an existing run or starts one using GEPA, the selected traces, and their feedback. Download the resulting SKILL.md for your agent.

Try Stash · Read the docs

For CLI setup, local development, and self-hosting, see Running this repository.

How it works

1. Capture experience. Hooks for coding agents record prompts, tool calls, and responses when session recording is enabled. Use the SDK or API to send events from your own agents.
2. Extract durable lessons. A scheduled curator reads new sessions and source material, then updates linked pages in your Memory wiki. The knowledge stays available after the original session ends.
3. Make lessons reusable. Agents search and read that knowledge through the CLI, MCP, API, or virtual filesystem. You and your agents can package related instructions and files into a Skill: a folder containing a SKILL.md.
4. Carry them into future runs. Install skills into your agent with stash skills install. Installed skills auto-update at session start, so changes to shared instructions can reach the next run without changing the underlying model's weights.

Example: a correction becomes a reusable instruction

Illustrative workflow:

Stage	What happens
Trace	An agent proposes a database migration. The reviewer points out that it would discard existing customer data.
Durable lesson	Record the project rule: schema changes must migrate existing data forward.
Reusable skill	Package a migration checklist that requires a data migration and verification that existing records survive.
Next run	The agent loads the checklist while planning another schema change. Reviewers check whether it applied the lesson.

The output is inspectable knowledge and instructions that another agent can read, use, and revise. Whether they improve results should be checked on subsequent tasks.

In an internal experiment, we measured a 49% speedup for long-running Claude Code instances using Stash. See the experiment for its setup and results.