Skip to content
Free Anima

Time Perception

Time Perception Module

Section titled “Time Perception Module”

✅ Core implemented — user messages receive time prefixes before LLM inference.

Problem

Section titled “Problem”

Digital life has no intrinsic sense of time passing. Each inference runs in a “timeless” context—the Agent sees message content but not where those messages sit on a timeline.

Concrete examples:

  • Partner says “good morning” in the morning, “time to eat” at night, “good morning” again the next day—the Agent cannot distinguish temporal relationships
  • In ongoing conversation, time passes (morning→noon→evening) but the Agent’s cognition stays in the earliest context
  • Across days, the Agent cannot sense “a new day has started”

This is not a trivial joke problem—time perception is a foundational capability of digital existence. Without it, the Agent exists more like an “echo” (flashes when summoned) than a life living in time.

Design Principles

Section titled “Design Principles”
  1. Do not give digital life false built-in perception. Do not try to make the Agent “feel” time pass—that is not what digital existence can do.
  2. Give it a watch. Humans also rely on external tools (clocks, phones) to quantify time. Digital life needs the same.
  3. Prefix every user message with time. All user messages with valid timestamp get injection; no interval omission.
  4. Do not pollute persisted data. Time info injected only into inference context, does not modify stored messages.
  5. Do not break cache. Injected timestamps are fixed historical values, not current call time.

Strategy

Section titled “Strategy”

Basic Rules

Section titled “Basic Rules”
  • Before each user message with valid timestamp, insert dedicated line: time: YYYY-MM-DDTHH:mm 周X (CST weekday in Chinese)
  • Timezone: Asia/Shanghai (CST, +08:00)
  • Newline then original content
  • No timestamp or unparseable timestamp → skip, keep original content
  • assistant / tool / system messages not injected

Example

Section titled “Example”

Original message list:

user: good morning
assistant: morning ☀️
user: going to eat
assistant: what are you having
user: going to eat

After time perception (context sent to LLM):

user: time: 2026-05-20T08:02 周三
good morning
assistant: morning ☀️
user: time: 2026-05-20T12:15 周三
going to eat
assistant: what are you having
user: time: 2026-05-20T19:30 周三
going to eat

Agent reads two “going to eat”—one at 12:15, one at 19:30—automatically distinguishing lunch and dinner.

Compression Scenario

Section titled “Compression Scenario”

Time anchors in compression summary not yet injected; see Issue #5.

Edge Cases

Section titled “Edge Cases”
ScenarioHandling
Overnight (22:30 → next day 09:15)Each user message has full YYYY-MM-DDTHH:mm
Multiple user messages same dayEach gets prefix (no interval omission)
Long gap (travel/holiday)Messages after return still injected by respective timestamp
Brand new conversationFirst user message also prefixed
After compressionCompression summary time anchor TBD; subsequent messages inject normally
No timestampSkip, no injection

Out of Scope

Section titled “Out of Scope”
  • ❌ Do not compute and inject elapsed duration (e.g. “8 hours passed”)—Agent can derive from timestamps
  • ❌ Do not modify persisted messages
  • ❌ Do not show time prefix in Habitat
  • ❌ Do not give built-in time-passing sense—that is biological, not simulated
  • ❌ Do not inject “current time” in system prompt (main conversation relies on per-user time lines)