Open any Anthropic community thread and you'll find some version of this question:

I had a long conversation with Claude three weeks ago about a project. Can I open it up and continue like nothing happened?

The short answer is no. The longer answer is more interesting: it's not that the technology can't do this. It's that no AI tool has built the infrastructure that would make it dependable.

Dependable long-term context is one of the most-requested AI features, and one of the hardest to deliver. Here's why it keeps failing and what it would take to actually solve.

What users mean by "resume"

When users ask if they can resume an old conversation, they're not asking to scroll back through chat history. They're asking for something specific:

The model should remember the substance of what was discussed
The model should remember decisions that were made
The model should remember the project's current state
The model should not need everything re-explained

In other words: pick up exactly where you left off, weeks later, with no setup.

This is different from "reading old chats." Reading old chats puts the cognitive burden on you to extract relevant context and feed it back to the model. Real resumption puts the burden on the system to maintain continuity.

Why current tools fail at this

Every major AI assistant has some flavor of memory. None of them do dependable long-term resumption. The reasons fall into four buckets:

Memory is opt-in, fuzzy, and per-fact

Claude's memory, ChatGPT's memory, Cursor's memory — they all work by extracting individual facts the model thinks are worth remembering. "User prefers TypeScript." "User is working on a billing system." "User dislikes verbose explanations."

This works for preferences. It fails for project state.

Project state isn't a list of facts. It's a tangle of decisions, dependencies, in-progress threads, and "we tried X, it didn't work, here's why." Per-fact memory can't capture this. By the time you've extracted enough discrete facts, you've also fragmented the context that made them coherent.

Old conversations aren't indexed for retrieval

Most AI tools store conversation history as a flat list. You can scroll through it. You can't query it by topic, time, or relevance.

When you start a new conversation about "the database migration," the system doesn't automatically pull in the three previous conversations where you discussed it. They exist somewhere in your history, but they're not surfaced.

Good retrieval is possible — semantic search over conversation history is well-understood — but most consumer AI tools haven't shipped it. Some power-user tools have. None feel seamless.

Context windows can't hold a month of work

Even if old conversations were retrievable, the model would have to fit them in its context window. A month of substantive work might be hundreds of thousands of tokens. Even Claude 4's 200K window only stretches so far.

This forces summarization. Summarization loses fidelity. The summary captures the high-level decisions but misses the nuance that often matters — "we considered X but rejected it because of Y" becomes "we did Z." The reasoning vanishes.

Larger context windows help but don't solve. You need a system that intelligently selects what to include, not just one that includes everything.

State drift

Even if you could resume perfectly, the world has moved on. The codebase changed. Your priorities shifted. The decisions made three weeks ago might no longer apply.

A naive resumption pulls in stale context. The model confidently references decisions that no longer hold. You spend the first ten minutes correcting it. This is worse than starting fresh because the model is confidently wrong.

Real resumption needs a mechanism to detect drift and trigger a reconciliation step. "Three weeks have passed since the last session on this topic. Here's a summary of what we'd discussed. Is anything no longer accurate?"

What dependable long-term memory would require

The technical pieces exist. They just haven't been assembled into a product yet:

Topic-anchored memory, not fact-anchored

Organize memories around topics — projects, ongoing threads, recurring themes — instead of extracted facts. When you start a conversation about "the database migration," the system pulls in the topic vault, not a list of disconnected facts.

Topic vaults can hold richer context: decisions, alternatives considered, current state, open questions. They can be structured (sections for decisions, status, history) instead of bullet points.

Semantic retrieval over conversation history

Index past conversations with embeddings. When a new conversation starts, automatically retrieve the most relevant previous sessions and offer to include them.

This doesn't mean stuffing everything into the context window. It means showing the user: "You discussed this topic in three previous sessions. Want me to summarize the relevant decisions?"

Hierarchical summarization

For long-running topics, store summaries at multiple levels:

Session-level: what happened in this conversation
Topic-level: the rolling state of this project
Decision-level: the key choices made and why

When you resume, the system pulls in the topic-level summary first. If you ask about a specific decision, it surfaces the decision-level detail. If you want the full history, it can drill into session-level transcripts.

Time-aware drift detection

Track when topics were last discussed. When you return to a topic after a long gap, the system explicitly surfaces this: "It's been three weeks since you worked on this. Here's where we left off. Anything to update?"

This turns drift from a problem into a feature — the gap becomes a reconciliation moment rather than a silent source of stale context.

User control over what carries forward

Not every old conversation is worth resuming. Users need an easy way to mark threads as closed, archived, or no longer relevant. The system should treat archived threads as historical record, not active context.

Conversely, users should be able to pin threads as "always relevant" — your master project documents, your standing preferences, your reference material.

Why the gap persists

If the pieces exist, why hasn't anyone shipped this?

Partly because retrieval over conversation history is messier than it sounds. Real conversations contain off-topic digressions, abandoned threads, corrections that overwrite earlier statements. Extracting clean state is hard.

Partly because the user interface is unsolved. "Here are five old conversations that might be relevant" is a worse experience than "the model just knows." The seams of the system show.

Partly because the major AI tools optimize for stateless interactions. ChatGPT, Claude, Cursor — they're built around the conversation as the unit. Long-term continuity is bolted on as memory features, not designed in as state management.

The tools that will solve this are the ones that treat memory as the primary abstraction, with conversations as the input that updates it. That's a different architecture than what most AI assistants ship today.

Until then

The practical workaround is to externalize state. Don't rely on the AI tool's memory. Maintain your own.

For each ongoing topic:

Keep a topic document (markdown file, Notion page, whatever you'll actually open) with current state
Update it at the end of each substantial session
Start new sessions by pasting it in

This is laborious. It's also the only thing that gives you dependable long-term context with current tools. Until AI assistants build the memory infrastructure users have been asking for, the user is the memory layer.

That won't last forever. But it's where we are right now.