Help your agents learn from experience.
Agents generate valuable experience every time they work: successful approaches,
failed attempts, and human corrections. Stash captures that history and makes
its lessons available to future runs.
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.
For CLI setup, local development, and self-hosting, see Running this repository.
How it works
- 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.
- 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.
- 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. - 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.
1 maintainer
Metadata
1 maintainer
Release files for stashai 0.1.426
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| stashai-0.1.426.tar.gz | 223.3 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| stashai-0.1.426-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 498.4 kB
Release files / stashai-0.1.426.tar.gz
| Download URL | stashai-0.1.426.tar.gz |
|---|---|
| Size | 223.3 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
a367325331b466e29df90159fcd513a0baf6ede546a2ad559461dde53c7418d5
|
|
BLAKE2b-256 checksum How to use checksums |
49204b9b1212e09a7a2731b5581d2f15423ae9a44b6e7cc8a6349454787bf6c2
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 9, 2026.
Transparency logRelease files / stashai-0.1.426-py3-none-any.whl
| Download URL | stashai-0.1.426-py3-none-any.whl |
|---|---|
| Size | 275.1 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
a85ba587c689ddb7dd4a97062b83fb953b4894cce807f60605a66d7440ba70fe
|
|
BLAKE2b-256 checksum How to use checksums |
c4179e72a7eab13bb9575aa750e28512e1bf37d54ae930d83b20d2f207b61180
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 9, 2026.
Transparency log
