When you use an AI assistant daily over months, something happens that most users do not think about until it goes wrong: the model accumulates context about you that you can no longer reconstruct yourself.
ChatGPT's memory feature knows the name of your startup, your role, your preferred communication style, conclusions from conversations six months ago. That knowledge exists inside OpenAI's system. You did not write it down anywhere else. If OpenAI changes its memory system, deprecates the feature, or your account is locked for any reason, that accumulated context is gone.
This is not a hypothetical risk. It is the default outcome of using hosted AI tools that control their own memory systems.
The lock-in structure of AI memory today
The AI memory landscape looks superficially competitive — ChatGPT memory, Claude Projects, Mem.ai, Notion AI, and a dozen others all claim to "remember" things for you. But the competition is at the user interface layer. At the data layer, your memory is stored in each vendor's proprietary system, accessible only through their interface, and non-portable in any meaningful sense.
A scan of what "export" means across the major platforms:
ChatGPT: The data export includes conversation history as JSON. The "memories" ChatGPT has stored about you — the distilled facts it uses to personalize responses — are accessible in Settings but not bulk-exportable in a format you can re-import anywhere. You can read them; you cannot take them with you.
Claude: Anthropic offers conversation history export as JSON. Claude Projects do not have a bulk export path for project knowledge. If you built up a Project for six months, you cannot move that knowledge to a different tool without manually reviewing and re-entering it.
Mem.ai: Mem.ai has a bulk export function. It outputs .mem files and CSV. Neither format is designed to be re-imported into a different knowledge management tool. Clean Markdown it is not.
Cognee: Self-hosted, so you own the data by definition. But there is no hosted product and no packaged export UX.
In each case, the data is technically yours in a legal sense. In a practical sense, it is not portable.
What HeurChain's vault export does
HeurChain stores every item you write to your tenant in a vault directory on the server. The vault contains your original text — not just the embedding vectors. This is a deliberate design decision. Embeddings without the original text are useless for portability.
From the dashboard, two buttons handle vault management:
Flush working memory to vault: Recent writes go into Redis first (fast working memory), and consolidate to the vault directory on a schedule. If you want to make sure everything is captured before downloading, flush first. The process takes under a minute.
Download vault: Downloads a .zip of your complete tenant memory as Markdown files, plus a manifest.json index. No waiting period. No request form. No 30-day GDPR process.
The output is clean Markdown. This is intentional. Markdown is readable without any software. It can be imported into Obsidian, indexed by any full-text search engine, passed to any AI tool as context, or migrated into a future memory system that does not exist yet.
Using the export with Obsidian
The vault download is designed to be Obsidian-compatible out of the box. Each memory item is exported as an individual .md file with YAML frontmatter:
---
id: mem_01hwz3q...
source: claude-session
stored_at: 2026-04-12T14:32:00Z
---
The user's preferred chunk size for retrieval is 512 tokens with 10% overlap,
based on empirical testing with bge-base-en-v1.5 in April 2026.To import into Obsidian: unzip the vault download into your Obsidian vault directory. Each memory item becomes an Obsidian note. Obsidian's built-in search and link graph work immediately — no plugin required.
From there, you can add bidirectional links, tag notes, or use any Obsidian plugin that operates on Markdown files. The memory you built up through months of AI agent interactions is now part of your personal knowledge base, readable and navigable without any AI involvement.
Why this matters for long-term AI use
The pattern of AI tool adoption tends to go: try it, rely on it, become dependent on the context it has accumulated, realize you can't leave because the context is trapped.
This is not malice. It is the natural result of memory systems designed to serve retention metrics rather than user autonomy. A memory system that is easy to export is a memory system that is easier to cancel. From the vendor's perspective, lock-in is a feature.
Our view is that sustainable AI infrastructure requires the opposite posture. If your memory is portable, you can use the best tool for each task without accumulating debt. You can migrate to a new tool without re-explaining yourself from scratch. You can shut down a subscription without losing months of accumulated context.
The 30-day export window after cancellation is in our Terms of Service as a hard commitment, not marketing copy. After 30 days, your data is permanently deleted. This is the correct default: your data should not persist on our servers indefinitely after you stop paying.
If you want AI memory you actually own, heurchain.com/pricing is where to start. The vault download is available on all plans from day one.