Introduction

Latticing is a Python library for Behavior Latticing — a technique for transforming raw interaction traces into a structured, multi-layered graph of behavioral insights.

Given a stream of user interactions (e.g., screen activity logs, chat histories, click traces), Lattice:

1. Observes — uses an LLM to infer what the user is thinking and feeling from their raw actions.
2. Generates insights — synthesizes observations into named, evidence-backed behavioral insights.
3. Builds layers — recursively merges per-session insights into cross-session patterns.
4. Visualizes — renders the resulting lattice as an interactive graph.

When to use Behavior Latticing

Behavior latticing is flexible across input type. Examples of input types are as follows:

• Messages with LLM-based chatbots (e.g., ChatGPT, Claude) (link)
• Interactions with coding agents (e.g., Claude Code)
• Screen recordings of computer usage (link)

See our Gallery for what others have built using Behavior Latticing.

Quick look

from lattice import Lattice, AsyncLLM, SyncLLM, Separator
import os

l = Lattice(
    name="Alice",
    interactions=interaction_traces,   # list of sessions
    description="the user's conversation history",
    model=AsyncLLM(name="claude-sonnet-4-6", api_key=os.getenv("ANTHROPIC_API_KEY")),
    evidence_model=AsyncLLM(name="claude-sonnet-4-6", api_key=os.getenv("ANTHROPIC_API_KEY")),
    format_model=SyncLLM(name="claude-sonnet-4-6", api_key=os.getenv("ANTHROPIC_API_KEY")),
)

config = {
    0: {"type": "session", "value": "1"},   # one group per session → L1 insights
    1: {"type": "session", "value": "10"},  # every 10 sessions → L2 synthesis
}

await l.build(config)
l.save("lattice.json")