since: temporal context for LLMs

LLMs have no sense of time. since fixes that.

pip install pysince
from since import Store, since_time

Zero dependencies. No AI calls. Works with any provider.

Before and after

Before - ask a vanilla LLM about past conversation:

> What did we talk about last time?
I don't have information about previous conversations.

After - with since:

> What did we talk about last time?
Welcome back! It's been 2 days since we last spoke.
We were debugging your auth flow - specifically the JWT expiry issue.

The model sees a timeline, not a flat list.

Related MCP server: Agent State MCP Server

Quick start

from since import Store, since_time
from openai import OpenAI

store = Store("~/.since/chat.db")
client = OpenAI()

@since_time(store=store, timezone="Asia/Kolkata")
def chat(messages):
    return client.chat.completions.create(model="gpt-4o", messages=messages)

resp = chat(messages=[{"role": "user", "content": "hello"}])
print(resp.choices[0].message.content)

The prompt sent to the model includes:

Now: Wed Jul 01, 02:36 AM (night)
Session: 4m · 3 messages
Gap: 2m between messages
Stale: "config.py" (read:config.py) invalidated, 14m old

The model sees when things happened, how long ago the last message was, and what context is stale.

Stale-file detection (for coding agents)

Files change between agent turns. since catches that by stamping file reads with mtime and content hash, then surfacing staleness when the file changes.

from since.stale_files import stamp_file_read, check_and_invalidate

# After reading a file
stamp_file_read("config.py", store, "session_1")

# Later, file changed externally
if check_and_invalidate("config.py", store, "session_1"):
    # Stale warning appears in the prompt tail automatically
    stamp_file_read("config.py", store, "session_1")

MCP server (for Claude Code, Cursor, etc.)

pysince-mcp

Exposes 4 tools for agent integration:

• stamp_file_read: stamp a file after reading it

• check_staleness: check if a stamped file has changed

• session_duration: how long the MCP server has been tracking file stamps

• invalidate_source: manually mark events stale

Note: session_duration reflects the age of file-read stamps, not conversation length (the MCP protocol has no access to chat history).

TTL system

Works with any provider

OpenAI, Anthropic, Gemini: @since_time detects the response shape automatically. Pass extract_reply= for anything else.

@since_time(store=store, extract_reply=lambda r: r.content[0].text)
def chat(messages):
    return anthropic.messages.create(model="claude-3-5-sonnet-20241022", messages=messages)

Requirements

• Python 3.10+

• Zero dependencies

Install

pip install pysince

The PyPI name is pysince (the since name was taken on PyPI). Import and repo are since.

Tests

pytest

45 tests, zero external services required.