A private atlas of your ChatGPT history
Turn your ChatGPT history into usable memory.
Find the questions you repeat, facts you changed, themes you left behind, and ideas that connect across years of conversations. Your export is analyzed in this browser, then becomes a searchable map with evidence behind every suggestion.
Analyze my ChatGPT ZIP10 question domains 4 memory states 2 model profiles 0 archive uploads
Why this exists
Your archive remembers more than any single chat can.
A long ChatGPT history contains decisions revisited under different wording, personal facts that later stopped being true, projects scattered across unrelated chats, and patterns that are almost impossible to see one conversation at a time.
Memory Map treats those conversations as evidence. It counts what can be counted exactly, uses local embeddings where meaning matters, and always keeps a route back to the sources.
Complete capability atlas
Everything the map can show you.
50 shipped capabilities across five routes. These are current product behaviors, not a roadmap.
Your questions, counted honestly
Fast local rules read the full parsed export first, so useful totals arrive even if the embedding model cannot load.
8 stops
Your questions, counted honestly
Fast local rules read the full parsed export first, so useful totals arrive even if the embedding model cannot load.
- History overview
Conversations, messages, your prompts, approximate words, active days, date range, and longest activity streak.
- Conversation depth
Median message depth plus short, medium, and deep conversation groups.
- Activity rhythms
Conversation volume by month, with source date range and peak-period evidence.
- Recurring terms
Frequently used meaningful terms that can surface themes worth revisiting.
- Exact repeated questions
Near-identical wording grouped with count, first ask, last ask, and source conversations.
- 10 overlapping question domains
Math, health, software, money, career, learning, creative work, relationships, travel, and planning—each with query count, conversation count, monthly history, and evidence.
- Likely typo signals
A deliberately small disclosed misspelling list plus repeated-word checks—not a grammar or writing-quality score.
- Possible thread changes
Low-overlap adjacent prompts flag conversations that may contain several tracks before deeper semantic segmentation runs.
Meaning, not just matching words
A pinned embedding model runs inside the browser to connect differently worded questions, conversations, facts, and strands.
9 stops
Meaning, not just matching words
A pinned embedding model runs inside the browser to connect differently worded questions, conversations, facts, and strands.
- Automatic model selection
Chooses a compact English profile or multilingual profile from the writing systems in your prompts; manual selection remains available.
- Visible model cost
Shows the pinned model, revision, selection reason, and approximate first-download size: about 90 MB for accelerated compact analysis, 24 MB for compact compatibility mode, or 135 MB multilingual.
- Accelerated browser inference
Uses WebGPU with larger bounded batches when the compact model and browser support it, then retries the complete ordered workload in portable WebAssembly mode when needed.
- Semantic topic clusters
Groups sampled conversations by meaning and names them against 20 broad topic anchors plus distinctive terms from your history.
- Interactive topic graph
Node size represents conversation count; edges represent semantic proximity; selecting a node opens its evidence.
- Semantic repeated questions
Finds similar intent even when the wording changed, while keeping these groups distinct from exact repeats.
- Four memory states
Current, updated, explicitly refuted, and possible contradiction—detected only from first-person statements, never assistant assertions.
- Memory change history
Keeps earlier and later wording, dates, update or refutation cues, semantic similarity, lexical overlap, confidence, and source conversations.
- Semantic conversation strands
Splits substantial multi-prompt conversations into ordered strands with visible boundaries, snippets, continuity, and confidence.
How your map changes over time
Monthly evidence distinguishes a momentary spike from a theme that is growing, fading, returning, or holding steady.
9 stops
How your map changes over time
Monthly evidence distinguishes a momentary spike from a theme that is growing, fading, returning, or holding steady.
- Topic evolution
Monthly histories and local trend labels for semantic topic clusters.
- Question-domain evolution
See whether health, software, planning, learning, and the other lenses become more or less prominent.
- Language-signal evolution
Tracks curiosity, frustration, urgency, uncertainty, excitement, appreciation, and neutral/direct wording by month.
- Five explicit trend states
Emerging, fading, resurfacing, steady, or not enough history—descriptions of the series, never explanations of why you changed.
- Positive, neutral, and negative wording
A compact lexical tone distribution and negative-wording rate, explicitly separated from personality or mental-health claims.
- Evidence-backed reflection questions
Open questions triggered by repeats, changed or refuted memories, possible contradictions, frustration-wording spikes, stale memories, dormant themes, recurring terms, and activity peaks.
- Six-view visual atlas
A coordinated activity calendar, question-mix chart, language matrix, repeat landscape, topic-momentum plot, and conversation-shape board, each with readable data and source evidence.
- One global view period
Switch between the recent 3, 6, 12, or 24 months and all history; chronological charts, query tone, and language signals stay aligned to the same window.
- Seven-slide Story mode
A keyboard-, swipe-, and touch-friendly presentation turns scale, cadence, question routes, repeats, language, and topic movement into an evidence-aware personal recap.
Search that explains itself
The derived index is useful as memory because search results and every inferred insight retain a route back to the conversation evidence.
12 stops
Search that explains itself
The derived index is useful as memory because search results and every inferred insight retain a route back to the conversation evidence.
- Natural-language memory search
Search conversations, individual questions, fact candidates, topics, and semantic strands with a hybrid semantic and lexical index.
- Visible graph traversal
Each memory question creates a temporary query node, highlights up to four topic routes, and shows the route from question to topic to evidence before an answer appears.
- Local memory conversation
An optional pinned 77M browser model turns retrieved evidence into a short conversational answer without receiving the full ZIP.
- Six-stop evidence boundary
The local generator receives at most six labelled excerpts and four recent chat turns, keeping context inspectable and browser work bounded.
- Citation grounding validator
Generated drafts must cite retrieved S1–S6 stops. An uncited draft is withheld while the source evidence remains available.
- Change-aware and repeat-aware retrieval
Questions about changed, updated, refuted, or repeated memories prioritize the matching structured ledger state instead of relying on similarity alone.
- Short conversational continuity
The panel retains only the four most recent turns for follow-ups; generated answers never become facts, graph edges, or saved memory.
- Typed, scored results
Each result identifies what it is, shows why it matched, includes a similarity score, and opens its topic or source.
- Evidence drawer
Topics, repeats, facts, domains, typo signals, strands, trends, language signals, and reflections all open supporting conversations.
- Adjustable evidence confidence
Exploratory, balanced, conservative, and custom thresholds refilter inferred repeats, memory changes, strand boundaries, and reflections without rereading the ZIP.
- Transparent sampling
The report states how many conversations, prompts, facts, and thread prompts were embedded versus the full parsed totals.
- Keyboard and non-graph access
A complete topic list mirrors the SVG graph; graph nodes, tabs, dialogs, search, and import controls support keyboard use and visible focus.
Private and recoverable by construction
There is no application server waiting for the archive. Parsing, analysis, search, and optional persistence happen on the device.
12 stops
Private and recoverable by construction
There is no application server waiting for the archive. Parsing, analysis, search, and optional persistence happen on the device.
- Original ZIP import
Accepts current ChatGPT exports containing one or more conversations*.json files; no manual unzipping required.
- Incremental split-file parsing
Reads large conversation chunks sequentially in a Web Worker instead of loading the whole archive into the page at once.
- Attachments stay unread
Binary attachments, chat.html, and unrelated export files are ignored.
- No archive upload
Conversation text is not sent to an application API, account system, analytics payload, or model endpoint.
- Deterministic fallback
Totals, activity, domains, exact repeats, wording signals, typo candidates, and initial thread candidates remain useful if embeddings fail.
- Nothing saved by default
The page does not persist the archive or derived index unless you explicitly choose “Keep on this device.”
- Device-local save, restore, and forget
An opt-in IndexedDB snapshot restores the derived report and search index; incompatible old snapshots are rejected and removed; Forget deletes the saved copy.
- Visible progress and recovery
Parse, statistics, model, embedding, and clustering stages report progress; analysis can be cancelled, errors explain recovery, and a new archive resets the report.
- Step-by-step wait estimates
A live six-stop route separates archive opening, parsing, initial insights, model preparation, semantic mapping, and final assembly, with elapsed time and a continuously revised remaining-time estimate.
- Complete execution receipt
After the map is ready, a distinct end-to-end timer includes parsing, model download and preparation, embeddings, and report assembly alongside initial-insight time, compute route, and preserved candidate coverage.
- Explicit chat-model lifecycle
The local generator loads only after consent, discloses its approximately 105 MB cached download and English focus, and can be unloaded independently at any time.
- Single-tab compute guard
A browser lock prevents duplicate heavyweight archive analysis across tabs, while reset, unload, errors, and page exit terminate model workers and release the lock.
The privacy architecture
Your conversations never need an application server.
The website is static. The analysis pipeline lives in your tab and its Web Worker.
- 1Your original ZIP
You choose it from your device. It is not submitted through a web form.
- 2Incremental browser parsing
Conversation JSON chunks are read one at a time; attachments remain ignored.
- 3Local rules + embeddings
Statistics run first. A pinned Hugging Face model file downloads separately, then embeds bounded text on your device.
- 4Your report and search index
Nothing persists unless you explicitly save the derived memory to this browser.
Archive text, parsed prompts, fact candidates, embeddings, report, and search index.
Pinned embedding-model files from Hugging Face. Your prompt text is not part of that request.
Archive upload endpoint, Memory Map account, cross-device sync, or server-side analysis.
Know before you import
Built for reflection. Not diagnosis.
Memory Map is a good fit if you want to…
- Find recurring questions and decisions across many chats.
- Build a searchable index of your own prompts, topics, facts, and strands.
- Review what you explicitly updated, rejected, or may have contradicted.
- See how interests and wording signals changed over time.
- Keep the archive on your device and inspect the evidence behind suggestions.
It is not currently the right tool if you need…
- Attachment, image, audio, or exported-memory-file analysis.
- A generative biography, automatic life summary, or personality profile.
- Medical, mental-health, or emotional diagnosis.
- Verified truth—the fact ledger contains candidates from your wording.
- Accounts, team collaboration, cross-device sync, or server backups.
Your ticket in
Download the ZIP directly from ChatGPT.
Keep it zipped. Memory Map knows how to find split conversation files and ignores the binary attachments.
- 1
Sign in to ChatGPT, open your profile menu, and choose Settings → Data controls.
- 2
Under Export data, choose Export, then Confirm export.
- 3
Open the email or SMS from OpenAI, download the ZIP, and return here without unzipping it.
Next stop
Your history is already there. Make it navigable.
No account. No archive upload. Nothing saved unless you ask.