openfossil 0.1.8

Semantic failure memory for AI agents

openfossil

Python SDK for FOSSIL — semantic failure memory for AI agents.

Install

# with local embedder (recommended, no API key needed)
pip install openfossil[local]

# core only — bring your own embedder or use remote API
pip install openfossil

# with OpenAI embedder
pip install openfossil[openai]

Quickstart

from fossil import Fossil, FailureType, ResolutionType, Severity, TaskDomain

# local SQLite store — zero config
fossil = Fossil()

# point at the live community API
fossil = Fossil(api_url="https://fossil-api.hello-76a.workers.dev")

# record a failure after it happens
fossil.record(
    situation="agent extracting JSON from LLM output that returned markdown-fenced text",
    failure_type=FailureType.FORMAT_FAILURE,
    failure="json.loads() raised JSONDecodeError — model wrapped output in ```json fences",
    severity=Severity.MAJOR,
    resolution_type=ResolutionType.INPUT_SANITIZATION,
    resolution="strip markdown fences with regex before parsing",
    framework="langchain",
    model="llama-3.3-70b-versatile",
    domain=TaskDomain.DATA_ANALYSIS,
    verified=True,
)

# search before a task step
results = fossil.search("parsing structured data from LLM text output")
for record, score in results:
    print(f"{score:.2f} — {record.resolution.description}")

CLI

# record a failure interactively
fossil record

# search for similar past failures
fossil search "JSON parsing failure"

# search against the live API
fossil search "JSON parsing failure" --api-url https://fossil-api.hello-76a.workers.dev

# list recent fossils
fossil list

# check connection
fossil ping

Inject into agent context

from fossil import Fossil, format_for_injection

fossil = Fossil()

def run_step(task: str, system_prompt: str) -> str:
    results = fossil.search(task, top_k=3, min_score=0.5)
    fossil_context = format_for_injection(results)
    augmented_prompt = system_prompt + "\n\n" + fossil_context
    # pass augmented_prompt to your LLM call
    ...

format_for_injection returns a string you drop directly into any system prompt or context window. No framework lock-in.

Remote store

from fossil import Fossil

fossil = Fossil(api_url="https://fossil-api.hello-76a.workers.dev")
fossil.record(...)
results = fossil.search(...)

No local embedder needed when using the remote store — embeddings run on the API side.

Custom embedder

from fossil import Fossil
from fossil.embedder import OpenAIEmbedder

embedder = OpenAIEmbedder(model_name="text-embedding-3-small", api_key="sk-...")
fossil = Fossil(embedder=embedder)

Custom store location

fossil = Fossil(db_path="/path/to/my/fossil.db")

Context manager

with Fossil() as fossil:
    fossil.record(...)
    results = fossil.search(...)

SDK API

Method	Description
fossil.record(...)	Record a failure + resolution
fossil.search(situation, top_k, min_score, domain)	Semantic similarity search
fossil.get(id)	Fetch a single record
fossil.delete(id)	Delete a record
fossil.list(limit, offset)	Paginate all records
fossil.count()	Total records in store

Failure types

misinterpretation hallucinated_tool format_failure context_loss infinite_loop premature_termination scope_creep ambiguity_paralysis tool_misuse adversarial_input compounding_error

Resolution types

prompt_change tool_fix retry human_override context_injection schema_correction step_decomposition input_sanitization

Running tests

cd packages/sdk-python
pip install -e ".[dev,local]"
pytest

First run downloads the embedding model (~80MB, cached after that).