Stash Agent Memory
Turn agent work history into searchable knowledge and reusable skills.
- Source repo
- Fergana-Labs/stash
- Stars
- ★ 337
- Last updated
- 6d ago
- License
- MIT
- Primary language
- Python
- FA score
- 52/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code (Partial support)
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- High · needs real infrastructure
- You'll need
- Typical use
- A development team wants to retain successful coding-agent approaches and human corrections for later tasks.
- Not a fit if
- Users who only need standalone prompts without retained session history
- Teams unwilling to send session data to a server
- Source review
- 52/100 · Major gaps
What does this agent do, and when should you use it?
Stash provides infrastructure for capturing agent experience, keeping durable knowledge, and reusing skills alongside existing agents and models. When session recording is enabled, coding-agent hooks capture prompts, tool calls, and responses; custom agents can send events through the Python SDK or REST API. A scheduled curator extracts lessons from new sessions and source material, updating linked pages in a Memory wiki. Agents and users can search or read that knowledge through the CLI, MCP, API, or virtual filesystem, then package related instructions and files into a Skill containing a SKILL.md. The repository provides a CLI, REST API, MCP, and Python SDK, and documents local development and self-hosting; deployment requirements are in the running guide.
With session recording enabled, hooks for coding agents record prompts, tool calls, and responses; custom agents can also send events through the SDK or REST API. A scheduled curator reads new sessions and source material, extracts durable lessons, and updates linked Memory wiki pages. The CLI, MCP, API, and virtual filesystem let users and agents search or read that knowledge. Related instructions and files can be packaged as a Skill containing SKILL.md. Run stash skills install to install skills; installed skills update automatically at session start so later runs can receive shared instruction changes.
- A development team wants to retain successful coding-agent approaches and human corrections for later tasks.
- An engineer maintaining a database wants to turn a review lesson about preserving existing data into a migration checklist skill.
- A developer building a custom agent wants to send session events through the Python SDK or REST API for curation.
- A user of MCP-enabled agents wants later runs to search Memory wiki pages and read relevant knowledge.
- A team wants to bundle project instructions and files into a Skill that refreshes at session start.
How do you install or deploy this agent?
The README points to a running guide for CLI setup, local development, and self-hosting, but the provided material does not include installation commands, service dependencies, credential requirements, or deployment steps. It is not enough to give a copyable install command or complete first-run configuration.
How do you use this agent?
Enable session recording to capture sessions through coding-agent hooks, or send events from a custom agent through the SDK or API. The provided material does not include SDK initialization, API endpoints, or authentication configuration. Install a Skill with:
stash skills installInstalled skills update automatically at session start. Specific CLI commands for searching knowledge and MCP configuration are not included in the provided material.
What are this agent's strengths and limitations?
- Covers a full experience-reuse path from session capture and scheduled curation to knowledge retrieval and skill installation.
- Offers several integration paths: Python SDK, REST API, MCP, CLI, and virtual filesystem.
- Skills are folders containing SKILL.md, so users and agents can inspect, use, and revise their contents.
- Installed skills refresh at session start, allowing shared instruction updates to reach later runs.
- The provided material omits deployment steps, runtime dependencies, authentication setup, and a first-call example; evaluation and rollout require additional documentation.
- Curation depends on a scheduled curator, but scheduling configuration, failure handling, and update latency are not specified here.
- Session capture includes prompts, tool calls, and responses, so teams need to assess their session-data collection and server-storage arrangements.
- The provided material does not state hosted pricing, model costs, or self-hosting resource requirements.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Stash Agent Memory This agent | 52 · Major gaps | CLIFreemium | ★ 337 | 6d ago | Python | — |
| pg-aiguide: PostgreSQL Expertise for AI Coding | 33 · Major gaps | CLIFree | ★ 1.9k | 11d ago | Python | ChatGPT · Codex · Claude Code |
| Haystack | 52 · Major gaps | Library / SDKFree + model costs | ★ 27k | today | Python | OpenAI API |
| CodeSeek | 48 · Major gaps | CLIFree + model costs | ★ 768 | 2mo ago | Rust | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The README describes recording prompts, tool calls, and responses, then having a scheduled curator read sessions and source material, making the main data flow reasonably clear (2). The supplied files do not show permission boundaries, confirmation before recording, or a complete sensitive-data retention and redaction policy, so least privilege, confirmation, and external effects are only partially addressed (1 each). A builder test shows remote image URLs are omitted from error logs, but this is a narrow safeguard, so sensitive-data handling scores 2. Nightly dependency audits cover several manifests, but retain an exception for an unfixed ecdsa advisory, yielding 2. Rollback is mentioned only in the prompt-optimization overview and is limited in scope (1). Publisher identity is unverified and the files do not establish verified source attribution (1).
The README's capture, curation, retrieval, and skill workflow broadly matches the CLI and MCP entry points in pyproject.toml (2). Multiple services and external Aspose, model, and network integrations create availability dependencies, while the supplied material shows no fallback strategy (1). The tests provide concrete but narrow evidence of HTTP error propagation, session cleanup in a finally path, and source-redacting logs (2).
The documentation targets developers and names SDK, REST, MCP, CLI, coding-agent hooks, and a self-hosting guide (2). It explains memory, skills, and reward-model uses and sets some boundaries, including that AI judgments are not customer-satisfaction measurements (2). Recording is said to occur when session recording is enabled, but the supplied material does not show precise trigger configuration (1). Several integration modes are named, but environmental requirements are not detailed here (1).
The README is organized around the workflow and links to SDK, deployment, and topic-specific documentation, though those linked files are not included in the supplied evidence (2). It points to installation and self-hosting instructions without giving the steps here (1). Package and command names are explicit in project metadata (2). The README includes a workflow example, interface descriptions, and some limits on AI assessment and prototypes (2); limitations are present but incomplete (2). The MIT license text is complete (3). A package version and automated bump workflow are present, but no conventional changelog is supplied (1). The security policy identifies reporting, scope, and response targets (2); unverified publisher identity earns no credit.
The product yields inspectable knowledge and SKILL.md files, and the README says later task outcomes should be checked (2). Turning operational experience into reusable knowledge has a clear potential benefit, but the claimed 49% speedup is an internal-experiment claim and the supplied material does not support generalizing it (2). The capture, curation, service, and model components add operational complexity, while cost-benefit is not quantified (1).
Some README feature descriptions correspond to concrete CLI metadata and builder-test behavior (2). The supplied tests cover a PPTX subsystem and do not independently verify the core memory or learning claims, so cross-source corroboration is limited (1). The README labels the workflow example illustrative and distinguishes AI judgments from customer satisfaction; the speedup is labeled an internal experiment, though claims still rely primarily on project-authored material (2).
- The README says prompts, tool calls, and responses can be recorded; confirm recording controls, access control, retention, and sensitive-content handling before deployment.
- The dependency audit explicitly exempts an ecdsa advisory with no fix available; review the affected dependency path and mitigations.
- The 49% speedup comes from an internal experiment and should not be treated as a performance guarantee for other tasks, models, or deployments.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
pr-815-images gh-attach assets
-
gh-attach-assets gh-attach assets