Agentome

Coordinate index layer for agent context — Agentome weighs, doesn't retrieve.

Agentome is a framework for treating AI agent context like a biological genome: compress, store, and emit a verdict on what to trust versus what to reread — instead of stuffing raw text into context windows.

This package is a thin identity wrapper around Helix Context, the reference implementation. pip install agentome pulls helix-context and re-exports the full public API.

Install

# Core (server + packet mode, no optional backends)
pip install agentome

# Recommended for daily use — launcher + tray + OTel + MCP + codec
pip install agentome[all]

Both commands install the engine under the hood. If you want the full deps list resolved, see the Full dependency matrix section below.

Quick Start

from agentome import (
    HelixContextManager, load_config,
    ContextItem, ContextPacket, RefreshTarget,
)
from agentome.context_packet import build_context_packet

config = load_config()
manager = HelixContextManager(config)

# Ingest — provenance (source_kind, volatility, last_verified_at)
# auto-populated from the file extension.
manager.ingest(open("src/main.py").read(),
               content_type="code",
               metadata={"path": "/repo/src/main.py"})

# Decoder path — full assembled context for a downstream LLM
window = manager.build_context("How does auth work?")
print(window.expressed_context)
print(window.context_health.resolution_confidence)  # 0.0-1.0

# Index path — agent-safe packet with verdict + refresh plan
packet = build_context_packet("How does auth work?",
                              task_type="edit",
                              genome=manager.genome)
for item in packet.verified:
    print(f"[{item.status}] {item.source_id}  truth={item.live_truth_score:.2f}")
for target in packet.refresh_targets:
    print(f"reread: {target.source_id}  reason={target.reason}")

What is an Agentome?

An Agentome is the complete set of compressed, structured knowledge that an AI agent carries as persistent memory, plus the pathway layer that tells the agent which parts of that memory it should trust versus go re-verify. The weighing is as load-bearing as the storage.

• Genes store compressed knowledge units (code, docs, conversations)
• Promoter tags enable fast retrieval by topic
• Provenance fields (source_kind, volatility_class, last_verified_at) drive task-sensitive freshness labeling
• Coordinate confidence measures whether retrieval landed in the right region, not just whether the words overlap
• Co-activation (harmonic_links) builds associative memory over time
• Replication grows the genome from every conversation

The concept is described in the research paper: The Agentome

Agentome vs Helix Context

	Agentome	Helix Context
What	The concept / framework	The implementation
Analogy	"Genome"	"Human Genome Project"
Package	pip install agentome	pip install helix-context
API	Re-exports Helix Context	The actual engine
Extras	Mirror helix-context's	[launcher], [codec], [mcp], etc.
Version policy	Locked to helix-context	Semver, beta-track

Use agentome in application code for the framework framing (from agentome import ...). Use helix-context in infrastructure code that you want to make obvious is wired to the engine. Both imports resolve to the same Python objects.

Full dependency matrix

Core pip install agentome installs:

Package	Version	Source	Why
helix-context	>=0.4.0b1	SwiftWing21/helix-context	The engine
fastapi	>=0.110	via helix-context	HTTP server
uvicorn	>=0.29	via helix-context	ASGI runtime
httpx	>=0.27	via helix-context	HTTP client (proxy + probes)
pydantic	>=2.6	via helix-context	Schema models

That's enough to run agentome (alias for helix), the HTTP server, and /context / /context/packet with a SQLite genome.

Optional extras

Every extra mirrors the corresponding helix-context extra. pip install agentome[X] == pip install helix-context[X] in dep resolution.

Extra	Pulls	Enables
accel	orjson	Faster JSON encode/decode
embeddings	numpy, sentence-transformers, torch (transitive)	SEMA 20D cold-tier retrieval
cpu	spacy	CpuTagger for entities (NER)
mcp	mcp>=1.0	Run python -m helix_context.mcp_server for Claude Code / Cursor / Claude Desktop integration
nli	torch, transformers	DeBERTa relation-graph NLI backend (standalone; embeddings pulls these transitively)
otel	opentelemetry-sdk + exporter + instrumentation	Metrics + traces to Grafana/Prometheus when HELIX_OTEL_ENABLED=1
launcher	jinja2, psutil	Supervisor launcher (helix-launcher CLI, :11438 dashboard)
launcher-native	jinja2, psutil, pywebview	Launcher with native window wrapper
launcher-tray	jinja2, psutil, pystray (LGPL-3), Pillow	Launcher with system tray icon
ast	tree-sitter + 4 language grammars	AST-aware code chunking
scorerift	scorerift	Divergence monitoring bridge
codec	headroom-ai[proxy,code]>=0.5.21	Recommended — CPU-resident semantic compression (Headroom by Tejas Chopra, Apache-2.0). Replaces character-level truncation in the expression pipeline. See NOTICE.
dev	pytest, pytest-asyncio	Contributor test suite
all	Every extra except dev, launcher-tray, launcher-native	The full feature surface

Non-pip runtime deps

Some features need infrastructure outside the Python dep graph. None of these are required to start the server.

Dependency	When	Install
Python 3.11+	always	python.org or your OS package manager
SQLite FTS5	retrieval (Tier 3)	Bundled with Python's sqlite3 on all mainstream builds. Verify with python -c "import sqlite3; sqlite3.connect(':memory:').execute('CREATE VIRTUAL TABLE t USING fts5(x)')"
Ollama	default [ribosome] backend = "ollama" + /v1/chat/completions answer generation	ollama.com — ships a local ollama serve on :11434
spaCy model	[cpu] extra for NER	python -m spacy download en_core_web_sm (after installing the cpu extra)
OTel collector	[otel] extra + HELIX_OTEL_ENABLED=1	Any OTLP/gRPC collector on HELIX_OTEL_ENDPOINT (default localhost:4317). Typically paired with Prometheus + Grafana.

Recommended install profiles

For different use cases, here's what to pull:

# Minimal — just serve /context and /context/packet with SQLite
pip install agentome

# Agent host (Claude Code, Cursor, MCP-aware tool)
pip install agentome[mcp,codec]

# Daily developer — everything you'd realistically use
pip install agentome[all,launcher]

# Observability stack
pip install agentome[all,otel,launcher]

# CI / bench / tests
pip install agentome[all,dev]

Headroom composition (the compression path)

Agentome's default install uses character-level truncation for gene content — works but leaves compression quality on the table. With pip install agentome[codec], the expression pipeline routes through Headroom (Tejas Chopra, chopratejas/headroom) instead:

• Kompress — ModernBERT ONNX-based semantic compression (fast CPU)
• CodeAwareCompressor — language-aware for code content
• LogCompressor / DiffCompressor — contextual compressors for structured text

Headroom is opt-in to keep the core Agentome install small. When installed, Agentome activates it transparently — no code changes. Composition rule: prefer Headroom when it's available, fall back to char truncation when not.

Full Documentation

See the Helix Context README for complete documentation including:

• Two product surfaces (/context decoder path + /context/packet agent-safe index path)
• Weighing layer (freshness × authority × specificity × coord confidence)
• Pipeline lane reference
• Delta-epsilon health monitoring
• Horizontal Gene Transfer (genome export / import)
• Continue IDE integration
• MCP tool surface
• ScoreRift divergence detection bridge

License

Apache 2.0