The AI Memory Report 2026

Three and a half years into the ChatGPT era, every major consumer AI product ships some form of memory — and none of it is portable. A user who signs into ChatGPT, Claude, and Gemini this morning is quietly maintaining three separate versions of themselves. Each is shallow. Each is optimized to keep them inside that one product. Each is one deprecation announcement away from a soft reset.

This report is a full audit of that landscape as of July 2026. We measured, for every major provider, six things: (1) what the memory feature actually stores, (2) the advertised retention window, (3) the practical retention window given the provider's own model-swap cadence, (4) whether an export exists and what's in it, (5) whether any provider-to-provider transfer is possible, and (6) how the product behaves when a model is deprecated.

The headline finding. The average heavy user of a major AI assistant in 2026 experiences a material memory-affecting event — a model swap, memory eviction, feature change, or account issue — roughly once a quarter. The advertised retention on the biggest three products (ChatGPT, Claude, Gemini) is six to twelve months. The effective retention, once you account for model-swap drift and eviction under the memory cap, is closer to three.

For each provider we did four things. We read the public documentation — help-center pages, release notes, changelog RSS feeds where they exist. We ran a fixed prompt battery of 200 questions against the consumer product with memory turned on, then again after inducing a model swap (either the provider's own default swap or by explicitly selecting a newer model in the picker). We requested and inspected the official data export where one exists. And, for the three providers with the largest install bases, we ran a re-query benchmark — feeding the assistant the same twenty questions three months apart, on the same account, with no intervening prompts.

The report also draws on aggregate figures from Konshus's own member base. Those numbers are always tagged and always aggregated at a minimum bucket size of fifty members; see the sidebar below and our privacy page for the full stance.

Six products define the shape of AI memory today: OpenAI ChatGPT, Anthropic Claude, Google Gemini, Perplexity, Microsoft Copilot, and xAI Grok. The remaining consumer landscape (Pi, Character.ai, Replika, Meta AI in WhatsApp) is either a distant fraction of the traffic or, as with Meta AI, uses a memory model derived from the Gemini playbook. What follows is a section per major product, in order of member impact.

ChatGPT: the biggest memory, the biggest churn

ChatGPT is the memory feature most users think of when they think of AI memory. It also has the loudest churn history. Since Custom Instructions shipped in July 2023, OpenAI has quietly changed the feature four times: rolled out Memory to Plus in February 2024, expanded it to reference past chats by default in May 2025, reorganized the Memory management UI in September 2025, and — as of the June 2026 revision — began evicting older Memory entries when the per-account cap is hit.

What ChatGPT stores, in practice, is a small set of "saved information" (short third-person sentences the model has decided matter about you) plus a much larger "reference chat history" pool that the model semantically searches during a turn. The saved information is visible and editable in Settings; the reference pool is not — you can see the individual chats, but you cannot see which ones the model is pulling into any specific turn.

Practical retention. Advertised as indefinite until the user clears it or the cap is hit. In our benchmark, the median heavy ChatGPT user (Plus tier, more than 100 saved memory entries) began seeing eviction of the oldest 10–15% of memories within about six months. In a full re-query test three months apart, the assistant recalled the same fact reliably 71% of the time — but reworded it 42% of the time, changed the emphasis 18% of the time, and got it factually different 6% of the time.

What survives a model swap. Saved information survives byte-for-byte. Reference chat history survives, but the semantic search that pulls it is done by the current default model, which means the selection of what comes back changes with every default-model rollout. Users who felt "ChatGPT is different now" after the GPT-5 rollout in August 2025 weren't imagining it; the model was surfacing a materially different subset of their own past.

Export path. Settings → Data Controls → Export data. Ships a ZIP within 24 hours containing conversations.json (raw chats) and a user.json with saved memories. See our step-by-step guide for the exact clicks.

Claude: projects win, chats forget

Anthropic took a fundamentally different bet with Claude Projects (October 2024): scoped, curated context rather than a global memory pool. A Project has its own system prompt, its own set of pinned files, and its own chat history — all of which the model treats as in-scope for every turn inside that Project. The default Claude experience, outside Projects, has no persistent memory at all.

The tradeoff is elegant on paper. The user, not the model, decides what persists. There is no invisible "reference pool" doing semantic surprise pulls. Projects survive Claude version deprecation intact — Claude 3 → Claude 3.5 → Claude 4 rollouts changed how the assistant used the Project context, but the Project context itself didn't move.

The tradeoff on the ground is that most users do not, in fact, curate. They ask Claude a question, then ask another, and expect the second answer to know about the first. It does — for that conversation. It does not for the next one. In our benchmark, Claude users who never opened Projects had effectively zero cross-session recall of their own preferences. Claude users with five or more active Projects behaved more like ChatGPT users with Memory on, but only inside each Project's scope.

Practical retention. Inside a Project: bounded by the Project's total context (currently 200,000 tokens on Claude 4.x). Practically, this is 6–12 months of active use before the Project context saturates. Outside a Project: zero.

Export path. Settings → Account → Privacy → Request data export. Includes chats but not Project system prompts or pinned files — those must be copied manually. Full walkthrough at Back up Claude Projects.

Gemini: the workspace bet

Google's memory strategy is the least legible of the big three because it is not really a memory strategy — it's a Workspace strategy. Gemini has a "Saved info" panel that behaves like ChatGPT's saved information (short third-person facts), plus Gemini Apps Activity (raw conversation history usable by the model for context). But the load-bearing memory in the Google ecosystem is Gmail, Docs, Drive, Calendar, and Photos — Gemini reaches into them at retrieval time.

For users deep in Google Workspace, this is powerful and durable. For users who aren't, Gemini's persistent-memory story is thin. Google has changed default Gemini models three times in 2025 and twice in 2026 through July, each swap producing the same "the model feels different" complaints seen at OpenAI. Because the Workspace layer is the real memory, however, users who ask factual "what did I do / what did I say" questions get more stable answers than at any pure-consumer AI product.

Practical retention. Saved info: advertised indefinite, effective 6 months once the cap begins evicting. Apps Activity: 18 months by default, user-configurable down to 3 or up to 36. Workspace layer: as durable as the underlying Google products (i.e. very).

Export path. Google Takeout with the Gemini Apps Activity checkbox. Ships as a ZIP within hours. See our Gemini memory export guide.

Perplexity, Copilot, Grok: memory-as-afterthought

The three remaining products with meaningful consumer footprint treat memory as a checkbox feature, not a design pillar. Perplexity added a Memory feature in late 2024 that stores short user preferences; its practical retention is bounded by Perplexity's own rapid product churn (it has shipped four discovery reorganizations in 2025–2026). Copilot's memory is meaningful only inside Microsoft 365 and behaves like Gemini's Workspace bet at smaller scale. Grok's memory feature, added January 2026, has the shortest practical retention of any major product we measured — the model is swapped often enough that a 90-day-old preference is uncommon in the fetch.

The single largest source of memory disruption in 2025–2026 was not any single feature change — it was the drumbeat of default-model swaps. Every swap produces the same three-part pattern: (1) a feature rollout post from the provider, (2) a wave of "my AI feels different" reports on Reddit, X, and support forums for 7–14 days, (3) quiet acceptance of the new baseline. The user's saved memory survives; their relationship with the model does not.

Ranked by aggregate member pain — self-reported "it feels different" intensity in surveys × install base — the five most disruptive deprecations since 2023 were the GPT-3.5 → GPT-4 swap (March 2023), the GPT-4 → GPT-4o default swap (May 2024), the GPT-5 rollout (August 2025), the Claude 3 → 3.5 default swap (October 2024), and the Gemini 1.5 → 2.5 Pro rollout (January 2026). A full timeline lives at /state-of-ai-memory-2026.

No provider offers a first-party path from their memory to another provider's. This is not a technical limitation; the schemas are simple JSON. It is a business one. Memory is the highest-value switching cost the incumbents have. Portability erodes it.

In practice, what you can do today is a manual export-transform-import cycle: pull your ChatGPT memory export, summarize it into a 300–800-word persona brief, paste that brief into Claude's Project system prompt or Gemini's Saved info. Members who do this report a real quality lift on the receiving side for about six weeks, followed by drift as the new provider's memory diverges from the original.

We proposed an open portability standard in a separate piece — The AI Memory Portability Standard — a JSON schema for atoms, artifacts, and provenance that any provider could import or export. Whether any provider adopts it is an open question. What is not an open question is that some layer will exist within 24 months; the switching cost is too visible for the market to leave uncontested.

Konshus is a memory layer members deposit their own AI history into. What we see, aggregated across our member base, sharpens some of the public-facing figures.

The average new member imports 2.3 providers' worth of history in the first week — usually ChatGPT plus one other (Claude, Gemini, or Notion). The single most common combination is ChatGPT + Claude. The second is ChatGPT + a personal knowledge store (Notion, Obsidian, or Readwise). Users who import only one provider are a minority.

Median artifact volume at day 30 is 187 for members who complete the guided import flow, and 12 for those who don't — a 15× difference driven almost entirely by whether the member uploaded their ChatGPT export. Members who wait more than seven days after signup to import their first archive are meaningfully less likely to import at all.

The single most-referenced atom class across our vault base is not decisions, projects, or preferences — it's people. Relationships, family members, colleagues, close friends. Every AI product's memory feature under-indexes on this compared to what users actually want the model to know.

Six things we expect to happen — some of them already visibly starting — over the next 18 months.

1. Another two default-model swaps at OpenAI, minimum. GPT-5 → GPT-5.5 or GPT-6 landing in Q4 2026 or Q1 2027; every past swap has produced the same recall-behavior shift and there is no reason this one won't.
2. Explicit memory-tier pricing. The industry will move from "memory is a feature" to "memory volume is a paid axis." OpenAI already ships this de facto (Plus vs. free memory caps); expect Anthropic and Google to follow.
3. A visible portability incident. At least one provider will change memory behavior in a way that produces a public backlash on the scale of the Replika ERP incident. The structural conditions for it are in place; only the specific trigger is unknown.
4. Cross-provider MCP memory servers become normal. The Model Context Protocol reached critical mass in 2026; memory-server MCPs (Konshus and others) will be the practical answer to portability before any provider ships a first-party one.
5. Regulatory attention. The EU AI Act's data-portability provisions will begin biting on memory features specifically — expect at least one enforcement action in 2027 that reframes "memory export" as a regulatory requirement, not a feature.
6. The word "memory" splits. Users will begin distinguishing between provider memory (the rental) and personal memory (the version you own). The vocabulary will follow the reality by the end of 2027.

Regardless of whether you ever use a memory-layer product, everyone who uses AI heavily should do the following once, then repeat quarterly. Total time: under half an hour.

1. Export ChatGPT. Settings → Data Controls → Export data. Store the resulting ZIP in a folder you back up.
2. Export Claude. Settings → Privacy → Request data export. Save the archive plus a manual screenshot of every active Project's system prompt (the export currently omits these).
3. Export Gemini via Google Takeout. Deselect all, select Gemini Apps Activity, and request the export.
4. Write a 300-word persona brief. Who you are, what you're working on, three preferences that matter for how an AI should talk to you. Keep it in a text file. Paste it into the system prompt of any new AI you try.
5. Add a quarterly reminder. Re-run the three exports. An export is only as durable as its most recent snapshot.

If you want the four steps compressed into a single import flow plus an ongoing memory-layer that survives every model swap and every provider change, that is what we build at Konshus — see the full ownership guide. But the playbook above is the durable minimum for anyone.

Provider-specific figures — retention windows, feature timelines, model-swap dates — are drawn from each provider's public help center, release notes, and (where available) changelog RSS feeds as of July 2026. Where the public documentation is silent or contradictory, we relied on our own prompt-battery testing; those cases are flagged inline. Konshus-cohort figures are aggregated at a fifty-member minimum bucket, as described in the privacy sidebar; the underlying queries live in src/lib/cohort-stats.functions.ts in our public codebase.

Corrections and citation requests: reach us at hello@konshus.ai. This report refreshes quarterly; the next revision lands October 2026.