From the team behind aidrop.it — one workspace to build, host, and keep changing your code.
Your AI assistant forgets because it never remembered. Models have no built-in long-term memory: they see only the current context window — the text of this conversation, right now. Close the session and everything is gone; let the conversation run long and old details get compacted away mid-session. "Memory" is something you have to add, and there are four ways to do it, from quick fixes to a real memory layer.
This is why ChatGPT re-asks what you told it last week, and why Claude Code "loses the thread" after a long run. It's not a bug or a dumb model — it's architecture.
Why AI forgets: context window ≠ memory
A context window is working memory — fast, complete, and strictly temporary, like RAM. Whatever isn't in the window doesn't exist for the model. Sessions don't persist it, other tools can't see it, and its size is hard-capped. Anything you want remembered must be re-supplied to every session, either by you or by a memory system. (For the full map of memory layers, see AI agent memory explained.)
The mid-session version: compaction
Long sessions hit the window limit, so tools like Claude Code compact: they replace older messages with a summary. Summaries keep the gist and drop specifics — the exact constraint you stated an hour ago, the file path, the "don't touch this" warning. If your agent suddenly repeats a mistake you already corrected, compaction probably ate the correction. Long contexts also burn your usage limits — retrieval beats re-reading.
The fixes, from quick to durable
Custom instructions / system prompts fix tone and standing preferences — a few hundred words, re-read every session. Project files like CLAUDE.md or AGENTS.md give coding agents per-repo context, versioned in git. Native tool memory (ChatGPT Memory) saves facts automatically, but inside one vendor's silo. Each step up is more durable — and each still stops at one tool's boundary.
The durable fix: a memory layer outside the tools
The endgame is memory that lives outside any single tool and connects to all of them over MCP. Sessions end, compaction runs, tools change; the memory stays.
A useful habit regardless of tooling: when an AI gets something wrong, save the correction somewhere persistent. Corrections are the exact map of what your assistant keeps forgetting.
FAQ
Why does ChatGPT forget what I told it? Each conversation is its own context window. ChatGPT's Memory feature saves some facts across chats, but it's selective, opaque, and only works inside ChatGPT.
Why does Claude Code lose context in long sessions? When the conversation outgrows the context window, it's compacted into a summary. Details not written to a file or memory before compaction are lost.
Do bigger context windows solve this? No. Bigger windows delay compaction but still end with the session, still cost tokens to refill, and still don't transfer between tools. Persistence requires storage outside the window — a memory layer.
One workspace to build, host, and keep changing your code
The repository, what the project knows, the rules a change has to follow, and the path to an address — kept together, and reachable by the coding agent your team already uses.