Semantic failure memory for AI agents.

---

Every AI agent is born knowing nothing.

It fails. You fix it. Next week, someone else's agent fails the exact same way. The week after, yours does too — different input, same root cause.

FOSSIL gives agents a memory for reasoning failures. Record what went wrong and how it was fixed. Search past failures before acting. Stop rediscovering the same mistakes.

---

How it works

Three operations.

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

fossil = Fossil()

# 1. 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() failed — 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,
)

# 2. search before a task step
results = fossil.search("parsing structured data from LLM text output")

# 3. inject into your agent's context
from fossil import format_for_injection
context = format_for_injection(results)
# drop context into your system prompt — done

---

Install

# Python SDK + CLI
pip install openfossil[local]

# JavaScript / TypeScript
npm install @openfossil/sdk

# MCP server (Claude Code, Codex CLI, Cursor, any MCP client)
npx @openfossil/mcp

[local] pulls in sentence-transformers for on-device embeddings. No API key. No data leaves your machine.

Or skip local setup entirely — point at the live community API:

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

---

CLI

fossil init                      # seed local store from community pool
fossil record                    # log a failure interactively
fossil search "JSON parse error" # search past failures
fossil list                      # browse your archive
fossil export --out backup.json  # export to JSON
fossil ping                      # check connection

---

MCP server

Works with Claude Code, Codex CLI, Cursor, OpenClaw, and any MCP-compatible agent:

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

Your agent can now call fossil_search before any step and fossil_record after any failure — no code changes to your agent.

---

Why not just use logging

Log aggregators tell you what happened. FOSSIL tells you what to do about it — by finding the time it happened before.

	Logging	FOSSIL
Unit of storage	Log line / trace	Reasoning failure + resolution
Search	Keyword / filter	Semantic similarity
Primary use	Debugging after the fact	Injection before acting
Shareable	Rarely	Yes, opt-in community pool

---

Failure taxonomy

FOSSIL classifies reasoning failures — not exceptions.

misinterpretation hallucinated_tool format_failure context_loss infinite_loop premature_termination scope_creep ambiguity_paralysis tool_misuse adversarial_input compounding_error

Each type implies a class of resolutions. Full spec in FOSSIL.md.

---

Integrations

Agent / Framework	Method	Guide
Claude Code	MCP server	docs
OpenAI Codex CLI	MCP server	docs
OpenClaw	MCP + ClawHub skill	docs
LangChain	Python SDK	docs
CrewAI	Python SDK	docs
Vercel AI SDK	JS/TS SDK	docs

---

Self-hosted

git clone https://github.com/heyrtl/fossil
cd fossil/self-hosted
cp .env.example .env
docker compose up db -d

PostgreSQL + pgvector. Point the SDK at it:

fossil = Fossil(api_url="http://your-server:8745")

---

Repo

packages/
  sdk-python/   — pip install openfossil
  sdk-js/       — npm install @openfossil/sdk
  core/         — TypeScript types (@openfossil/core)
  mcp/          — MCP server (@openfossil/mcp)
apps/
  api/          — REST API (Cloudflare Workers + D1 + Workers AI)
examples/
  groq-agent/   — working demo, Groq free tier
schema/
  fossil.v1.json
skills/
  openclaw/     — ClawHub skill
docs/           — full documentation
FOSSIL.md       — protocol spec

---

Status

Protocol v1.0 stable. Python SDK, JS SDK, MCP server, REST API — all live.

Community API: https://fossil-api.hello-76a.workers.dev — no key required.

---

Contributing

Implementations in other languages are welcome. Conformant = validates against schema/fossil.v1.json. Read FOSSIL.md.

---

MIT