How do I give ChatGPT or Claude persistent memory across sessions?
You cannot make the model itself remember. Each session starts from the same fixed weights, so persistent memory has to live somewhere outside the model and be handed back to it at the start of the next task. Every setup that works, from a single instructions file to a memory server, is really the same two parts: a durable store, and a rule for what gets loaded back in.
Nothing in the model changes when you explain something
Correcting an AI in a chat does not retrain the model or change its weights. Some AI products keep a separate memory outside the model, but that memory remains tied to that product. If you want durable context that can survive across sessions, people, and supported AI clients, the durable state has to live somewhere outside the conversation itself.
So the question is not how to make the AI remember. It is where the context lives between sessions, and what decides which parts of it come back.
Every setup that works is the same two parts
A folder of Markdown, an instructions file the AI loads every session, a handoff summary pasted into the next chat, a small memory server, a vector store. They look like different answers and they are not: each one is a place to keep the context, plus a rule for what gets handed back. The interesting differences are in the second part.
If what you want is a comparison of those setups and whether to assemble one yourself, that is its own question and it has its own page: do I need to build my own AI memory system?
An instructions file is not a workaround
For a set of rules you want applied every single time, a file the AI reads at the start of the session is the correct answer, not a stopgap. It is visible, it costs nothing to load, and you can tell exactly why the AI did what it did. Nothing below is an argument for replacing it.
The reason it holds is that every rule in the file is a rule you want applied every time. Everything that follows is about what happens when that stops being true.
Four things that appear once the memory has a job
None of these are storage problems, which is why adding a bigger store does not fix them.
Two notes disagree
You corrected a rule in March and again in July. Both lines are in the file. Nothing in the file says which one is current, so the AI picks one.
Loading all of it stops helping
Past a certain size, handing over everything buries the three rules that mattered. The problem moves from remembering to choosing.
Someone else needs part of it
A teammate needs your deployment rules but not your notes on their performance review. A file is all or nothing.
The memory cannot do anything
A better answer is where the file ends. Nobody wanted a better answer. They wanted the ticket filed and the record updated.
Durable context, corrected as you work, resolved for the next task
HexaHQ keeps the context outside any one chat and outside any one AI. When you correct something while working, your AI can write the corrected version to HexaHQ as knowledge, and you review what it saved. On a later task, the parts that apply to that task come back with it. Not all of it, and not because a similarity score was high, but because that knowledge is attached to the work it governs.
Two things this is not. HexaHQ does not copy your conversations, and it does not reach inside the model. What persists is what was deliberately saved: facts, rules, corrections, preferences, and reusable steps, with a version history you can read.
The other half is that the context can lead to work getting done. Once a rule is in HexaHQ, the same rule that shapes the AI's decision also travels with it into the systems where the work actually lands.
Your teammate gets the rule, not your notes
Context in HexaHQ carries an audience. A note can stay personal to you, go to a few named people, go to your whole organization, or go to one external collaborator. That same boundary decides what a person's AI may use on their behalf, so your teammate's AI gets the deployment rule and never sees the rest of your notes.
A correction on Monday, a filed ticket three weeks later
The point of the example is the last step. The context is only worth keeping if it changes what gets done.
-
The AI gets it wrong
A support lead asks their AI to write up a customer escalation. It grades the severity as routine.
-
The person corrects it, once
"Escalations from customers on the enterprise plan start one level higher." The AI writes the corrected rule to HexaHQ. The lead checks what it wrote.
-
Three weeks later, a new chat
Different session, different day. The lead asks for the week's escalations to be triaged. The rule arrives with the task.
-
The work gets done
Two escalations are graded correctly and filed in Linear under the lead's own account, after the lead approves the write.
Questions people ask next
How do I stop Claude from starting from scratch every session?
Put the context it needs somewhere outside the chat and make sure it is handed back on the next task. An instructions file does this for a fixed set of rules. HexaHQ does it for context that keeps changing, because your AI can write corrections to it while you work and read back the parts that apply next time.
How do I keep AI from forgetting project context?
Separate the two problems. Storing the context is the easy half. The harder half is deciding which parts of it belong in front of the AI for the task at hand, which is what stops a growing file from making answers worse.
How do I persist context between AI chats?
Through a store both chats can reach. HexaHQ is that store for durable rules, facts, and reusable steps, and the same material is available whether the next chat is in ChatGPT, Codex, Claude Chat, Claude Code, or Claude Cowork.
Is this the same as the memory feature in ChatGPT or Claude?
Native memory can be useful for personal continuity inside one product. HexaHQ is for the organizational context you want to own, review, scope, and reuse across supported AI clients.
Does HexaHQ store my chat history?
No. It stores what was deliberately saved as knowledge, a workflow, or a capability, and a record of the actions it ran. It does not copy transcripts and it has no access to anything inside the model.
Do I have to give up my Markdown files?
No. Files, repositories, and documents stay useful as source material. What HexaHQ adds is a single current version of the rules your AI should apply, with the audience and the scope attached.
Related pages
Stop re-explaining the same context.
Save what you correct while you work, and start the next task from it. Free to start.