Tempo Tokens

Tempo tokens are time-grounded embeddings or annotations that represent temporal structure — explicitly encoding timing, recurrence, and duration into a model’s input at inference time.

Time-Aware Transformers

Transformers today use fixed or learned positional encodings that reflect the order of tokens during training. That works well for language, but it doesn’t capture semantic time—like the difference between Monday 3PM and Saturday 10AM.
There’s currently no built-in way to represent:

• Recurring events

• Temporal gaps or proximity

• Day-of-week or time-of-day patterns

• Future planning vs past reasoning

Language Models Lack Clock and Calendar Awareness
While LLMs can talk about time, they don’t use time as part of their internal context. They don’t know when something happened. They don’t model how long real-world events take. And they have no sense of what’s coming next unless it’s explicitly mentioned in a prompt.

This limits their ability to plan, coordinate, or help users manage real-time tasks. By grounding model inputs in structured time—especially using external sources like .ics calendar data—we can start to bridge this gap and build agents that better understand the temporal context their users inhabit.

Proposed Function of Tempo Tokens

At training or inference, tempo tokens can be inserted as special annotations that;

• Encode absolute time: @2025-04-30T14:00Z
• Encode relative time: @+3d, @past-1w
• Encode calendar structure: @Monday, @weekend, @EoQ
• Encode rhythmic recurrence: @biweekly, @every-Tuesday

These could be actual tokens (<tempo:mon_14_00>) or embedded features added via a side-channel (like extra inputs in fine-tuning or adapters).

Enables user adaptability and specification for agent tasks

Training-Time Augmentation

Use .ics or calendar data to:

• Inject time-structured context into dialogue or task completion datasets.

• Align text spans with event timing and sequence (e.g., doctor notes, project logs).

• Generate training examples where temporal cues improve performance on downstream tasks (planning, scheduling, coordination, summarization).

Example

Transform

Submit payroll to HR next Tuesday May 13th. 

into:

"Submit payroll to HR. <tempo:due-2025-05-01> <tempo:project-deadline>"

Inference-Time Augmentation

At inference:

• Inject known temporal signals from .ics context into the prompt to improve grounding

• Align responses with upcoming events, time constraints, or temporal reasoning goals

• Enable planning agents to prioritize actions in realistic and user-defined time horizons

Use Cases with .ics Integration

Chronologue Command Agent

• Converts .ics entries into tempo tokens for prompt construction
• Use upcoming event timing to suggest actions, order and organize tasks, detect and manage conflicts in scheduling

Reasoning and CoT

• Tempo tokens give context to memory traces: when it happened, how often, what comes next

Time-Aware Retrieval and Summarization

• “Summarize the last week’s events” —> uses tempo-annotated memory traces to construct a coherent and verifiable narrative

Feasibility:

• Can be done without architecture changes using adapter layers or embedding projection (to link events in chronological time)
• Using time as an index dimension

Logistically:

• .ics data is structured and timestamped, making it the ideal import format for tempo token generation — assembling largescale datasets and synthetic datasets
• Many existing datasets (project logs, Github commits, emails) could be aligned with time

Showcasing how an .ics calendar event can be parsed into tempo tokens — these tokens can be retrieved to augment inference-time system-level prompts for temporally grounded chain-of-thought (CoT) reasoning.

1. Sample .ics Event

BEGIN:VEVENT
UID:shipment-arrival-20250430T150000Z@chronologue.ai
DTSTAMP:20250429T120000Z
DTSTART:2025-04-30T15:00:00Z
DTEND:2025-04-30T15:30:00Z
SUMMARY:Inbound Shipment — Med Supplies (Batch #23B)
LOCATION:Receiving Dock 3
DESCRIPTION:Urgent delivery. Requires cold chain transfer to Storage A by 16:00.
END:VEVENT

2. Structured Event JSON

{
  "uid": "shipment-arrival-20250430T150000Z@logistics.ai",
  "title": "Inbound Shipment — Med Supplies (Batch #23B)",
  "start": "2025-04-30T15:00:00Z",
  "end": "2025-04-30T15:30:00Z",
  "location": "Receiving Dock 3",
  "description": "Urgent delivery. Requires cold chain transfer to Storage A by 16:00."
}

3. Tempo Tokens for Inference Context

<tempo:2025-04-30T15:00Z>
<tempo:Wednesday>
<tempo:Afternoon>
<tempo:shipment>
<tempo:urgent>
<tempo:deadline-2025-04-30T16:00Z>
<tempo:location-ReceivingDock3>
<tempo:handoff-required>

These tokens give the agent:

• Exact time of arrival

• Deadline for transfer

• Event type (shipment)

• Urgency and task completion constraints

• Location grounding

4. Injected CoT Prompt for Temporal Reasoning

Prompt:

You are coordinating logistics for medical shipments. Use the schedule below to assign responsibilities and detect any conflicts or risks.

System Prompt

Upcoming Event:
Inbound Shipment — Med Supplies (Batch #23B)
<tempo:2025-04-30T15:00Z> <tempo:Wednesday> <tempo:Afternoon>
<tempo:urgent> <tempo:deadline-2025-04-30T16:00Z> <tempo:handoff-required>
<tempo:location-ReceivingDock3>

Current cold chain staff availability:
- Luis (shift ends at 15:15)
- Priya (available 14:00–18:00)
- Forklift 1 in use until 14:55

Recent delivery delays: 12–18 min average at Receiving Dock 3

Simulated Chain-of-Thought Output

The shipment is scheduled for 15:00 with a cold chain deadline at 16:00. Luis’s shift ends too early to ensure delivery, so Priya should be assigned. There’s a risk of a 12–18 minute delay at Receiving Dock 3, which means backup cold chain gear should be pre-cooled in Storage A. Forklift 1 will be available in time for transfer. Assign Priya and notify Receiving Dock to prioritize unloading.

​

Inference Benefits

Tempo tokens allow the model to:

• Predictively allocate resources based on event timing

• Reason about deadlines, urgency, and gaps

• Align agents (human or AI) with scheduled activity

• Anticipate failure modes (e.g., delays, shift coverage, equipment availability)

Required Fields

Optional Fields