Cross-session memory is the ability of an AI agent to remember context, decisions and conversation history across separate sessions, so it never starts from scratch.
If your chatbot forgets everything you told it yesterday, it doesn't have cross-session memory — it has a context window and nothing else.
This guide covers what cross-session memory is, why most chatbots forget between sessions, how it works under the hood, and exactly how we run it in production — 20+ scheduled agents a day, on one Dell laptop, sharing a single memory store.
Cross-session memory is the ability of an AI system to carry information from one session to the next.
A session is one continuous conversation with a deadline attached to it.
When the session ends, the context window empties.
An agent with cross-session memory writes the important parts somewhere durable before that happens.
Next time you open a new session, the agent searches that durable store and picks up where you left off.
It's the difference between a temp who needs the whole briefing again every morning and a colleague who already knows your customers, your prices and your pet peeves.
Most chatbots are stateless by design.
Every new session opens with a fresh, empty context window.
Nothing from the previous session is loaded, because nothing was saved.
Default ChatGPT, most website chatbot widgets and most free AI tools work this way.
They are built for one-shot questions, not for ongoing work.
That's why you tell a chatbot your business name, your opening hours and your pricing — and it asks you again the next day.
It's not a bug in the model.
It's a missing storage layer.
The model can remember perfectly well inside a session.
The problem is the session ends and everything evaporates.
Under the hood, cross-session memory is a three-step loop.
That's it.
No retraining, no fine-tuning, no vector database required for most use cases.
In Hermes Agent, the store is a SQLite session database with FTS5 full-text search.
Every session is logged automatically, and a session_search tool retrieves past conversations by keyword, date or session ID.
On top of that sits a permanent memory system: short personal notes that survive every session and can only be overwritten deliberately.
That two-layer design — searchable history plus a curated memory file — is the pattern that works in production.
Want this running in your business without learning the plumbing?
See AI Suite setup options →Kits from £497 one-off · 30-day money-back guarantee
The context window is short-term working memory.
It lives only during one session and has a hard token limit.
A 200k-token context window is impressive — and it still forgets everything the moment the session closes.
Memory is the long-term store that survives sessions.
Think of the context window as your desk and memory as the filing cabinet.
You can only hold so much on the desk at once.
But if you file things properly, you can pull any document back out tomorrow.
The most common mistake in agent design is trying to solve memory problems with a bigger context window.
Bigger desk, same filing system.
What you actually need is the filing cabinet.
We run cross-session memory in production every day at AI Suite.
Here is the real setup, not a vendor demo:
None of this needed a vector database, a GPU or a big cloud bill.
It needed a memory layer that actually persists.
The hard part was never the model — it was remembering what the model had already been told.
A support agent that remembers the customer's history answers in one step instead of five.
"I emailed you last week about my invoice" becomes actionable instantly.
Your agent knows which leads were emailed, which replied, and what the last message said.
No duplicate outreach, no dropped threads.
An email agent with persistent memory never re-asks for context it already has.
It categorises, drafts replies and remembers preferences across weeks.
Scheduled agents research in one session, draft in the next and publish in a third — all consistent, because they share memory.
A central agent holds your SOPs, prices and customer facts so every other agent answers from the same source of truth.
That is the difference between agents that feel like a team and agents that feel like a series of cold calls.
We build agents with shared memory for UK businesses — set up for you, live in days.
See the Call & Booking Agent →Or start free with a 30-minute strategy session
With open-source agent frameworks, cross-session memory costs nothing extra.
It is a built-in feature of Hermes Agent, not a paid add-on.
Your real costs are the agent build itself and the model API usage.
At AI Suite:
Compare that to the cost of a customer repeating themselves, a lead going cold, or a follow-up missed.
Memory isn't a luxury feature.
It's the thing that makes automation feel intelligent instead of robotic.
It is the ability of an AI agent to remember context, decisions and conversation history across separate sessions, so it doesn't start from scratch every time. It's the difference between a chatbot that asks the same questions daily and an agent that behaves like a colleague who already knows your business.
Most chatbots are stateless: each session opens with a fresh, empty context window. Anything not written to a persistent store is lost when the session ends. That's why default ChatGPT and most website chatbots forget everything you told them yesterday.
The agent logs every conversation, decision and tool output to a persistent store. On a new session it searches that store and injects the relevant history into the context window. In Hermes Agent this is a SQLite-backed session database with FTS5 full-text search, plus a curated memory file for permanent facts.
The context window is short-term working memory that exists only during one session and has a token limit. Memory is the long-term store that survives sessions. A 200k context window still forgets everything when the session closes unless key facts are written to memory first.
With open-source frameworks like Hermes Agent it costs nothing extra — it's a built-in feature. Managed AI setup from AI Suite starts at £497 one-off, and a full multi-agent system with shared memory runs from £199/month.
Hermes Agent has it built in (memory store, session database and searchable history). ChatGPT, Claude and Gemini have limited native memory features, while agent frameworks like n8n and custom Hermes builds add persistent memory to any chatbot. We run 20+ scheduled agents daily on one Dell laptop, all sharing one memory store.
It depends on where the memory lives. Local memory stores (like Hermes Agent's default setup) never leave your machine. Cloud options keep data on the provider's servers. We recommend local-first memory for sensitive business data.
Stop re-explaining your business to your own software.
Get an agent with cross-session memory — built for your workflow, live in days.
Browse the store → Free: 21-point AI checklist →Last updated: 2026 · AI Suite, UK · Cross-session memory for AI agents and chatbots