How DriftOS Works

DriftOS organizes conversation history into semantic branches. Only the relevant branch goes to the LLM.

The Problem with Traditional Memory

Traditional chatbot memory sends the entire conversation history to the LLM on every request:

User: "Tell me about dogs"
User: "What about black holes?"
User: "I want to plan a vacation"
User: "Back to dogs—what's the adoption fee?"

Traditional approach: All 4 messages + responses = 3,891 tokens sent to LLM

The LLM has to process irrelevant context (black holes, vacation) to answer a question about dogs.

DriftOS routes each message to a semantic branch. When you ask about adoption fees, the LLM only sees the "dogs" branch:

Branch 1 (dogs):         2 messages → 847 tokens
Branch 2 (black holes):  1 message → pruned
Branch 3 (vacation):     1 message → pruned

Result: 78% token reduction. Faster responses. Focused context.

DriftOS analyzes each message and makes one of three decisions:

The message continues the current conversation branch.

Example:

Current Branch: dogs
User Message: "What's the adoption fee?"
Decision: STAY (continue in dogs branch)

The message introduces a new semantic topic. DriftOS creates a new branch.

Example:

Current Branch: dogs
User Message: "Tell me about black holes"
Decision: BRANCH (create new branch: black-holes)

The message returns to a previous topic. DriftOS switches to an existing branch.

Example:

Current Branch: black-holes
User Message: "Back to the dog thing—what's the fee?"
Decision: ROUTE (switch to existing branch: dogs)

Note: ROUTE is only available with DriftOS Core (LLM routing). DriftOS Embed creates new branches instead.

DriftOS offers two routing strategies:

Uses Groq-powered LLMs to analyze semantic meaning.

Strengths:

Performance:

Best for: Multi-topic conversations where accuracy matters.

Uses vector similarity to compare message embeddings.

Strengths:

Performance:

Best for: High-volume applications where speed matters.

Tradeoff: Cannot ROUTE to old topics—creates new branches instead.