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
