since: temporal context for LLMs

since gives anything in an LLM's context a sense of how old it is — conversation turns, file reads, tool outputs. One library, zero dependencies.

pip install pysince
from since import Store, since_time

---

For chat apps

Wrap your chat function with @since_time. Every message gets a timestamp. The model sees a timeline instead of a flat list.

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)

Before: ask a vanilla model about past conversations. It has no memory.

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

After: the model sees when each message happened and how long the gaps were.

> 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 prompt tail the model sees:

Now: Wed Jul 01, 02:36 AM (night)
Session: 9h 2m · 4m active · 3 sittings · 8 messages
Gap: 6h between messages
Stale: "config.py" (read:config.py) invalidated, 14m old

The model knows when things happened, how long ago, and what context is stale.

For coding agents (MCP server)

Same primitive, aimed at files. Stamp a file when you read it. Check staleness before editing.

pysince-mcp

stamp_file_read — call after reading any file you intend to edit:

Stamped read: read:/path/to/config.json

check_staleness — call before editing a previously-read file:

Stale=True (content changed, mtime changed) read 4m ago

If the file changed, the agent re-reads it before acting on cached content. No daemon, no polling — just mtime and content hash comparison at the next turn.

Setup: your MCP client needs a trigger line telling the agent when to call the tools. For Claude Code or Cursor, add to your system instructions:

For every file you read, call stamp_file_read immediately. Before any edit, call check_staleness on files involved in the change.

TTL system

Class	Decay	Use case
permanent	Never	Facts, identity
slow	Session age	Normal conversation
event	On invalidate()	File reads, tool outputs
ephemeral	5 minutes	"ok", "thanks"

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.