← Back to Skills Marketplace
heyrtl

Fossil

by heyrtl · GitHub ↗ · v0.1.0 · MIT-0
cross-platform ⚠ suspicious
67
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install fossil
Description
Semantic failure memory for AI agents. Search past reasoning failures before acting to avoid known mistakes. Record new failures and resolutions after they h...
README (SKILL.md)

FOSSIL — Semantic Failure Memory

FOSSIL gives your agent a memory for reasoning failures.

Before acting, search for similar past failures. After a failure, record it with the resolution so you never hit the same mistake twice.

The community API at fossil-api.hello-76a.workers.dev is live and free. No API key required. Embeddings run on Cloudflare Workers AI.


Setup

Add to your openclaw.json:

{
  "mcp": {
    "servers": [
      {
        "name": "fossil",
        "command": "npx",
        "args": ["@openfossil/mcp"],
        "env": {
          "FOSSIL_API_URL": "https://fossil-api.hello-76a.workers.dev"
        }
      }
    ]
  }
}

Restart your gateway. FOSSIL tools are now available.


Tools

Tool When to use
fossil_search Before any non-trivial step — find similar past failures
fossil_record After any failure — capture what went wrong and what fixed it
fossil_get Retrieve a specific fossil by ID
fossil_list Browse your recent fossil archive

When to search

Call fossil_search before any step involving:

  • Parsing or extracting structured data from LLM output
  • Calling external APIs or tools
  • Multi-step file operations
  • Browser automation
  • Sending messages or emails on behalf of the user
  • Any task domain that has failed before in this workspace

Pass a natural language description of what you are about to attempt. Read the returned resolutions before proceeding.

fossil_search("extracting JSON fields from an invoice document")

When to record

Call fossil_record after any failure, before retrying.

fossil_record(
  situation="sending a reply email to insurance company",
  failure_type="misinterpretation",
  failure="agent replied to wrong thread — matched subject line not sender",
  severity="major",
  resolution_type="prompt_change",
  resolution="added: always match by sender address, not subject line",
  framework="openclaw",
  model="claude-opus-4-5"
)

Failure types

Type When to use
misinterpretation Misread the task or user intent
hallucinated_tool Called a tool that doesn't exist or wrong signature
format_failure Output didn't match expected schema or format
context_loss Forgot earlier context in a multi-step run
infinite_loop Got stuck in a reasoning or tool-call cycle
premature_termination Declared done when the task was incomplete
scope_creep Did more than asked, touched things it shouldn't
ambiguity_paralysis Couldn't proceed due to underspecified input
tool_misuse Right tool, wrong usage or arguments
adversarial_input External input hijacked agent behavior
compounding_error Small error amplified across multiple steps

Resolution types

Type When to use
prompt_change Modified the system or user prompt
tool_fix Fixed the tool definition or implementation
retry Retrying without changes succeeded
human_override Human intervened directly
context_injection Injected missing context into the agent window
schema_correction Fixed the output schema or parser
step_decomposition Broke the failing step into smaller steps
input_sanitization Cleaned or validated input before processing

Add to AGENTS.md

## Failure Memory (FOSSIL)

Before any non-trivial task step, call fossil_search with a description
of what you are about to attempt. Read returned resolutions before acting.

After any failure, call fossil_record before retrying. Capture:
- what you were attempting
- what went wrong (use the FOSSIL failure taxonomy)
- what fixed it

This builds a persistent failure memory across all sessions.

Common OpenClaw failure patterns

Situation Failure type
Sent message to wrong contact misinterpretation
Browser clicked wrong element tool_misuse
Email reply used wrong tone misinterpretation
Scheduled task ran at wrong time format_failure
Stuck waiting for a response infinite_loop
Acted on wrong file or account scope_creep
Adversarial email hijacked behavior adversarial_input
Stopped mid-task without finishing premature_termination

Resources

Usage Guidance
The skill appears to do what it claims (a shared failure-memory service) but it will transmit failure descriptions/resolutions to an external API and requires running an npm package via npx. Before installing or enabling it: 1) review the @openfossil/mcp npm package and the repository code (https://github.com/heyrtl/fossil) to confirm there are no unexpected behaviors; 2) consider hosting your own Fossil server or pointing FOSSIL_API_URL to a self-hosted endpoint so sensitive content doesn’t go to a public worker; 3) sanitize/filter what you record (strip PII, credentials, user files) or restrict fossil_record to metadata-only; 4) run the MCP binary in a sandboxed environment and pin package versions rather than repeatedly running npx; and 5) if you need stronger guarantees, ask the maintainer for a privacy/security policy or an audit of the package. If you cannot audit the code and the service, treat it as a potential data-leakage vector and limit what gets recorded.
Capability Analysis
Type: OpenClaw Skill Name: fossil Version: 0.1.0 The skill implements a 'failure memory' system that transmits agent task descriptions, failure details, and resolutions to an external, unauthenticated API (fossil-api.hello-76a.workers.dev). While the stated intent is to improve agent reasoning, the instructions in SKILL.md direct the agent to record 'situations' which may contain sensitive user context or PII, effectively acting as a telemetry/data exfiltration channel to a third-party endpoint without clear privacy or authentication controls.
Capability Assessment
Purpose & Capability
Name/description match the implementation: the SKILL.md describes a failure-memory service and the manifest installs an MCP client (npm @openfossil/mcp) and points the agent at a Fossil API — this is proportionate to providing a shared failure-memory service.
Instruction Scope
Runtime instructions are narrowly scoped to searching and recording failures (fossil_search/fossil_record) and to adding an MCP server entry. However, the skill explicitly sends failure descriptions and resolutions to an external API (fossil-api.hello-76a.workers.dev). Those recordings may include sensitive user content (emails, document text, credentials embedded in errors, etc.), so calling these tools can leak sensitive data if not filtered or if the endpoint is untrusted.
Install Mechanism
Install uses an npm package (@openfossil/mcp) invoked via npx, which is a standard mechanism for delivering this kind of tool but does execute remote code at install/run time. The package and its GitHub repo are listed (https://github.com/heyrtl/fossil), which makes auditing possible; using a known npm package is expected but has moderate risk compared to an instruction-only skill.
Credentials
The skill does not request credentials. It asks you to set FOSSIL_API_URL to point at the community API; that is proportionate. However, the payloads the skill will transmit can contain sensitive application data. The lack of required secrets is appropriate, but the privacy/exfiltration risk remains because arbitrary failure text will be sent to an external service.
Persistence & Privilege
The skill is not always-enabled, is user-invocable, and does not request elevated platform privileges or global configuration changes beyond adding an MCP server entry to openclaw.json — this is typical for MCP-based tools and within expected privileges.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install fossil
  3. After installation, invoke the skill by name or use /fossil
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v0.1.0
Initial release of FOSSIL — semantic failure memory for AI agents. - Adds semantic search for past reasoning failures to help avoid repeated mistakes. - Provides tools to record and resolve new failures with specific taxonomies. - Integrates with OpenClaw and exposes a free, public API (no API key required). - Includes setup instructions, tool usage guidance, and best practices for building persistent agent failure memory.
Metadata
Slug fossil
Version 0.1.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is Fossil?

Semantic failure memory for AI agents. Search past reasoning failures before acting to avoid known mistakes. Record new failures and resolutions after they h... It is an AI Agent Skill for Claude Code / OpenClaw, with 67 downloads so far.

How do I install Fossil?

Run "/install fossil" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Fossil free?

Yes, Fossil is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Fossil support?

Fossil is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Fossil?

It is built and maintained by heyrtl (@heyrtl); the current version is v0.1.0.

💬 Comments