axiom-trace 1.3.0

Local-first, append-only trace vault for AI agents with cryptographic integrity

Axiom Trace

Local-first, append-only trace vault for AI agents with cryptographic integrity and hybrid search.

Installation

pip install axiom-trace

Quick Start

from axiom_trace import AxiomTrace

# Create a vault (defaults to .axiom_trace/ in current directory)
with AxiomTrace() as trace:
    # Record an event
    frame_id = trace.record({
        "event_type": "thought",
        "content": {
            "text": "Analyzing user request...",
            "rationale_summary": "Breaking down the problem"
        },
        "metadata": {"model_name": "gpt-4"}
    })
    
    # Query the vault
    results = trace.query("user request", limit=5)
    
    # Verify integrity
    status = trace.verify_integrity()
    print(f"Integrity OK: {status['ok']}")

---

Memvid Cloud Integration (Optional)

Axiom Trace uses Memvid for hybrid semantic search. You can enhance search quality with a Memvid API key.

Setup in Your Project

1. Create a .env file in your project root:

# In your project directory (e.g., my_project/)
echo "MEMVID_API_KEY=mv2_your_api_key_here" > .env

# Add to .gitignore to keep it secure
echo ".env" >> .gitignore

2. That's it! Axiom Trace automatically loads the key:

from axiom_trace import AxiomTrace

# Automatically uses MEMVID_API_KEY from your project's .env
trace = AxiomTrace()

Alternative: Pass Key Directly

trace = AxiomTrace(memvid_api_key="mv2_your_api_key_here")

How It Works

your_project/
├── .env                 ← Put MEMVID_API_KEY here
├── .gitignore          ← Add .env to keep key secure
├── main.py             ← from axiom_trace import AxiomTrace
└── .axiom_trace/       ← Default vault location

When you import axiom_trace, it uses python-dotenv to load your project's .env file automatically.

---

Observer Pattern

from axiom_trace import AxiomTrace, session, observe, set_global_trace

trace = AxiomTrace()
set_global_trace(trace)

# Context manager for sessions
with session(session_id="my-session") as obs:
    obs.record_user_input("What is the weather?")
    obs.record_thought("Need to call weather API", "Determining tool to use")
    obs.record_tool_call("weather_api", {"city": "NYC"})
    obs.record_tool_output("weather_api", {"temp": 72, "condition": "sunny"})
    obs.record_final_result("The weather in NYC is 72°F and sunny!")

# Decorator for automatic tracing
@observe(event_type="tool_call")
def search_database(query: str) -> list:
    return ["result1", "result2"]

---

CLI

# Record an event from JSON file
axiom record --vault ./my_vault --event event.json

# Query the vault
axiom query --vault ./my_vault --prompt "weather" --limit 5

# Export a session to Markdown
axiom export --vault ./my_vault --session <uuid> --out session.md

# Verify vault integrity
axiom verify --vault ./my_vault

# Show statistics
axiom stats --vault ./my_vault

---

Event Types

Type	Description
thought	Agent reasoning (requires rationale_summary)
tool_call	Tool invocation (requires tool_name in metadata)
tool_output	Tool result (requires tool_name in metadata)
user_input	User message
final_result	Final response
system_event	System notifications
error	Error with stack trace

---

Vault Structure

.axiom_trace/              # Default location (or specify your own)
├── vault.manifest.json    # Metadata + head hash
├── vault.lock             # Write lock
├── vault.mv2              # Memvid index
├── frames.jsonl           # Append-only frame storage
└── axiom.log              # Internal logs

---

Integrity Verification

Axiom Trace uses SHA-256 hash chains for tamper detection:

result = trace.verify_integrity()
# {
#     "ok": True,
#     "checked_frames": 150,
#     "head_hash": "abc123...",
#     "error": None
# }

Detects: modified content, deleted frames, reordered frames, manifest tampering.

---

API Reference

AxiomTrace(vault_dir=None, memvid_api_key=None, ...)

Parameter	Default	Description
vault_dir	.axiom_trace/	Vault storage location
memvid_api_key	From .env	Memvid API key for cloud features
redaction_enabled	True	Auto-redact secrets
auto_flush	True	Background flush every 5s

Methods

• record(event) → str (frame_id)
• query(prompt, limit=5, filters=None) → list[dict]
• export_session(session_id, out_path, format="md")
• verify_integrity() → dict
• stats() → dict
• close()

---

License

MIT