Docs · Compare
Every AI lab and every memory company now keeps something about you. What differs is what is kept, how it is boiled down, how it reaches the model, and whose it is. Read as of 13 October 2026.
| Product | What it keeps | How it is distilled | How it reaches the model | Whose it is |
|---|---|---|---|---|
| ChatGPT | A memory summary synthesised from all your chats; past chats, files and apps as sources | In the background, automatically, rewritten as facts change | Loaded into every chat; the summary is “partial by design” and editable | OpenAI’s app only; delete clears the summary, not the chats |
| Claude | An editable summary of standing facts, rebuilt every day; past-chat search on demand | Daily synthesis; import from ChatGPT or Gemini by pasting an export | The summary, plus a search when it would help | Anthropic’s app |
| Gemini | Memory learned from past chats, plus the information you save | Accumulates on its own; the mechanism is not published | Personalisation inside Google’s apps | Google only; off for Workspace, minors, the EEA and the UK |
| Meta Muse | Dossiers on you and on every person you mention, updated hourly; a nightly analysis of the day | Implicit and continuous; infers goals you never stated | Loaded into the agent’s own machine | Meta’s agent; exportable files; “forget” does not erase the source |
| Mem0 | Facts extracted from each exchange; an optional graph of entities | On write, a model decides to add, update, delete or leave each fact against its nearest neighbours | Searched at question time | The developer who integrates it; the person has no product |
| Supermemory | Source chunks, extracted facts in a graph, and a profile | “Dreaming”: grouped documents are merged and organised by a memory model | Searched at question time, plus the profile as standing context | Developer infrastructure; began as a consumer second brain |
| Honcho | A “representation” of each person, derived in the background from every message | A background process, with summaries every few dozen messages | An agent the app consults about the user | Developer infrastructure for app builders |
| Engram | A learned memory artifact trained from a company’s documents | Studied in advance and compressed; stale when the documents change | In the model’s context as trained memory, not retrieved text | The enterprise |
| Zep / Graphiti | A knowledge graph of conversations and business data that knows when each fact held | Facts get validity periods and are invalidated, not deleted | Graph search at question time | Developer infrastructure |
| Letta | Core memory blocks in context; archival and recall memory outside it | A background agent rewrites the blocks between turns | The blocks are always in context; the rest is searched | The developer’s agent |
| MemoryPlugin | Memory snippets in buckets, plus full chat history | None by design: duplicates merged and stale facts flagged, each fix approved by you | Injected into each chat by a browser extension or a connector | The person, across many tools |
| Nuramem | Records you wrote, kept by kind — decisions, plans, people, references, learnings, state — saved when you say so | Consolidated in the background into one compact picture, kind by kind | Loaded at the start of every chat as prior knowledge, in every AI and app you use | Yours: one-call export, deletion that deletes; the record your rules read |
Sources, read 13 October 2026
The Mem0 paper (arXiv 2504.19413) · Supermemory, how it works · the Honcho architecture · The Deep Feed on Engram and on the June memory wave · ChatGPT’s memory announcement of 4 June and its Memory FAQ · Claude’s memory guide, May 2026 · Gemini memory, July 2026 · Letta, sleep-time compute · Zep / Graphiti (arXiv 2501.13956) · TIME on Muse, 6 October · Muse Wonder’s Active Memory, 12 October · MemoryPlugin’s own pages. Each vendor’s figures are its own; none is independently verified. Something here out of date? Tell us.