LLM Chat History and Memory: What Carries Into a New Session?
A saved conversation, the context of an answer, and memory for later sessions are different things. Learn what carries forward and how to preserve the work you need.
Updated 5 min read
LLM chat history and memory are different layers: history stores exchanges, while memory is saved information an app can bring into a later response. What carries into a new session depends on what the app supplies, retrieves, or reads, rather than simply whether an old conversation still exists.
An LLM, or large language model, generates text from the information available to it. The surrounding app manages conversations and connections. Tools like fenbs keep an explicit work record outside a chat app, so connected assistants can read current decisions and progress without assuming they remember an earlier exchange.
What does a saved conversation actually preserve?
A transcript preserves what was said: your questions, the replies, and any other material the app includes in that record. It can be useful evidence of an approval or a source for reconstructing a task. Its existence alone does not establish that a different conversation has received those messages.
Keep four layers separate when deciding what the next session needs. Each answers a different question, and a system can have one without providing all the others.
| Layer | What it does | What it does not establish |
|---|---|---|
| Stored transcript | Keeps the exchanges you can revisit. | That a new answer has access to every exchange. |
| Active context | Provides information for the current response. | That the same information will appear in a later session. |
| Retrieval | Fetches relevant saved information when requested. | That the result is complete, current, or approved. |
| Durable work record | Keeps explicit decisions, task progress, and next steps. | That an assistant read it before acting. |
Use the transcript to establish what happened in a conversation. Use a current project record to establish what should happen next. The AI memory hub covers the wider family of memory questions.
What is active context, and why can it change?
Active context is the information available while the model produces an answer. It can include your latest message, earlier messages, instructions, documents, and tool results. The context window is the limited capacity available for that material and the generated response, measured in tokens, the small units used to process text.
Anthropic’s context window documentation (opens in a new tab) distinguishes this working material from the data used to train a model. It also explains that chat interfaces can manage a growing conversation by changing which material remains in the window. A long saved transcript therefore need not match the input of every later response.
Reading a document successfully today does not prove the assistant will receive it tomorrow. Ask whether the next session will reopen the source, receive a summary, or start without it. Our explanation of what an LLM is provides the model background; the practical check here is which inputs support this answer.
How does retrieval bring older information back?
Retrieval means fetching saved information to help answer a request. Instead of placing every old conversation into every response, an app can search for a relevant exchange and supply what it finds. The retrieved text becomes input the model can use now.
For a concrete example, Claude’s chat search and memory guide (opens in a new tab) describes searching earlier conversations and linking back to them through citations. The same guide provides controls for reviewing remembered information. These are app capabilities around the model, rather than proof that every earlier message is always present.
A search result still needs interpretation. Say a design discussion contains three proposed headlines and one final approval. Finding a proposed headline establishes that someone considered it; it does not establish that the team selected it. Ask for the source of the approved choice and check the surrounding exchange.
What can persist between separate sessions?
Apps can carry information forward through saved conversation state, stored memory, documents, or connected work tools. OpenAI’s conversation state documentation (opens in a new tab) explains that individual text generation requests are stateless, meaning they do not independently retain the previous request. It also describes ways to supply earlier messages or persist a conversation for later use.
The distinction is between the model generating an answer and the system supplying its inputs. A new session need not start without context, but continuity needs a mechanism. Ask what was stored, how it reaches the next response, and whether your new app can access that source.
For one app’s specific controls, see ChatGPT memory versus chat history; for the broader categories of persistent information, see agent memory. Neither a convincing recollection nor a familiar writing style proves that an assistant has the current project record.
Which details deserve a durable work record?
Preserve information that changes the next action. Keep the current decision separate from brainstorming, and distinguish a finished draft from a completed review. A useful record is small enough for someone to inspect and explicit enough to correct.
- Decision: what a person approved, who chose it, and the work it applies to.
- Current state: what changed and what remains unfinished.
- Evidence: what was checked, the result, and any remaining uncertainty.
- Next step: where to resume and what is waiting on a person.
In the headline example, save the approved wording and the pending review task rather than only a summary saying the campaign is progressing. When a person changes the choice, record the replacement clearly. fenbs keeps these distinctions in Decisions and rules, tasks, and Where we left off; how it works explains the starting context connected assistants read.
Keep the work record easy to check
Save the approved choice and the unfinished task in fenbs, where History records changes with who made them. Ask the next connected assistant to read that record and identify its starting point. You can then check the source of its answer instead of relying on a claim that it remembers.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How can you check what carried into a new session?
Start a fresh conversation and ask the assistant to read the saved record, state the current decision, and name the next unfinished step. Compare the answer with the source yourself. Testing inside the original conversation only shows that the assistant can use information available there.
If the answer is wrong, locate the missing link before adding more notes. Was the fact saved, was the right source available, did the assistant read it, and was it current? Correct the source or the starting instruction. Repeating an outdated fact in another place makes the next check harder.
How to do this with fenbs
- Create a free account, free for one person with no credit card, and add the project you want to continue across sessions.
- Review the useful outcomes of your conversation. Record choices people made in Decisions and rules, then add unfinished work as tasks on the kanban board.
- Write Where we left off with what was completed, what was checked, and the next step. Keep unresolved questions explicit.
- Connect a supported assistant using Integrations. The connection uses MCP, a standard way for an AI app to use an outside tool, with browser sign-in.
- Ask the connected assistant to read the project context before working and save progress before stopping. Verify the next session against the same record.
fenbs does not import ChatGPT or Claude chats, and it does not read their chat history. It is not the right fit if your only need is reopening an old exchange. Use it when reviewed decisions, progress, and an attributed history of changes need to remain available beyond the conversation.