Your AI agent ran a task yesterday. Today it starts from zero. How are you handling cross-session state?

[oliviacraft]

The problem

Your AI agent ran a research task yesterday. Compiled sources. Made decisions. Built context over 2+ hours.

Today you open a new session. Ask for a follow-up. It starts from scratch. No memory of what it found. No delta. No trend. Just the same work done twice, billed twice.

---

This is not a model problem. It is an architecture problem.

The agent is stateless by design. Your job is to give it a state layer — and most setups do not have one.

---

What I have seen people try

• Dumping long summaries at the start of each session (manual, breaks down fast)
• Saving raw logs and pasting them back (noisy, unstructured)
• Writing a memory.md file manually after each session (unsustainable)
• Using embeddings / vector DBs (overkill for most use cases)
• Ignoring the problem until drift becomes visible (the most common approach)

---

What actually works (in my experience)

The lightest-weight solution that actually sticks:

1. A structured MEMORY.md file with three layers: daily notes, extracted facts, and patterns
2. A heartbeat cron that runs at a fixed interval and updates those layers
3. A startup hook that loads the right context into the agent before it runs

This is the architecture behind the Personal Agent Starter Kit — a $17 setup that wires all three layers together with AGENTS.md, MEMORY.md, and a working heartbeat loop.

---

My question

How are you handling this in your own setups?

• Are you using a file-based memory layer, a vector DB, or something else?
• How do you decide what survives from one session to the next?
• Have you found a way to make the agent write its own memory reliably?

Curious what patterns are actually working — especially for agents running on a schedule (not interactive sessions).

---

If you are just starting out: the free starter kit includes a basic AGENTS.md + memory scaffold you can drop into any project.