Konshus.ai

Annual report · 12 min read · June 2026

State of AI Memory 2026

Every major event that has reshaped what your AI remembers about you, from the Replika wipe of February 2023 through the GPT-5 default swap and the rise of MCP. Fifteen disruptions, one underlying pattern, and a clear-eyed look at where this is going.

Timeline visualization of AI memory disruptions 2023-2026

Key finding

In the 40 months between February 2023 and June 2026, every major AI provider has shipped at least one event that materially affected what users' assistants remembered or how they behaved. The cadence is accelerating, not slowing. The only AI memory that has been continuously stable across this period is the kind that lives outside any one provider's app.

The timeline

  1. Feb 2023

    Replika

    Disabled ERP after Italy's Garante intervention; underlying model behavior shifted.

    Thousands of users described losing the personality of their AI companion overnight.

  2. Mar 2023

    OpenAI

    Brief outage exposed some users' chat titles to other accounts.

    First mainstream moment of doubt about ChatGPT memory privacy.

  3. Mar 2023

    OpenAI

    GPT-3.5 → GPT-4 rollout; behavior changed noticeably in many prompt patterns.

    Established the pattern of 'model release = personality reset.'

  4. Jul 2023

    OpenAI

    Custom Instructions feature shipped — first persistent persona setting.

    Made persistent context official but limited to two text fields.

  5. Nov 2023

    Character.ai

    First major filter tightening; many users reported personality changes.

    Set the precedent for ongoing filter-driven character drift.

  6. Feb 2024

    OpenAI

    ChatGPT Memory feature began rolling out to Plus users.

    First mass-market 'real' persistent memory; capped from day one.

  7. May 2024

    OpenAI

    GPT-4o launched, replacing GPT-4 as default.

    Recall behavior shifted; some users felt Memory had been 'reset.'

  8. Oct 2024

    Anthropic

    Claude Projects shipped.

    Brought scoped, curated context to Claude — distinct from ChatGPT Memory model.

  9. Apr 2025

    Rewind AI

    Sunsetting of consumer Rewind product announced.

    Demonstrated that even purpose-built memory products can disappear.

  10. May 2025

    OpenAI

    Memory expanded to reference past chats by default for Plus.

    Made the feature more powerful but also more opaque about what was being used.

  11. Aug 2025

    OpenAI

    GPT-5 launched as new default; older models began deprecation runway.

    Triggered the largest wave of 'my AI feels different' reports to date.

  12. Sep 2025

    Anthropic

    Claude 3 deprecated; Claude 3.5 became baseline.

    Projects survived; conversational habits did not.

  13. Jan 2026

    Google

    Gemini 1.5 deprecation; Gemini 2.5 Pro became default.

    Context behavior changed even though saved info nominally persisted.

  14. Mar 2026

    Multiple

    MCP (Model Context Protocol) reached critical mass across Claude, ChatGPT custom GPTs, and third-party clients.

    First real cross-provider standard for context — created the technical opening for portable memory layers.

  15. Jun 2026

    Where we are

    Every major provider has shipped some form of memory; none of it is portable; the next model release is always one quarter away.

    The structural problem is unchanged. Continuity remains the user's job.

The four patterns that repeat

Across all fifteen events, the same four patterns recur. Anyone designing their relationship with AI memory long-term should plan around these, not against them.

  1. Model swaps are silent. Behavior changes the day a new default ships, but the change is rarely announced in user-facing terms. The data is usually still there; the way the model uses it is not.
  2. Filter changes feel like memory loss. When a provider tightens what its model will say (Replika, Character.ai, ChatGPT moderation passes), users experience it as the AI forgetting who they are together — even when nothing was technically deleted.
  3. Deprecation has no rollback. When an old model is sunset, the specific personality tuned to that model is gone. Custom GPTs, Claude Projects, and saved memories all survive in form but not in feel.
  4. Continuity is the user's job. No provider has ever made memory their headline feature, because their incentive is to keep you inside their app, not portable. Continuity survives only when the user — or a tool acting on the user's behalf — keeps memory outside the providers' walls.

What changes in 2026

Two structural shifts have opened genuine new ground this year. First, MCP (Model Context Protocol) has reached enough adoption to function as a real cross-provider context standard — Claude, ChatGPT custom GPTs, Cursor, and an expanding set of third-party clients all speak it. For the first time, a memory layer can plug into multiple AIs without per-provider integration work.

Second, dedicated memory tools are no longer novel. A small category of products — Konshus among them — is designed from day one around model-swap survival, full portability, and member-controlled deletion. That category did not exist as a recognized class in 2023.

Net: the tools to solve this have arrived, even as the underlying incentive structure has not changed. Users who actively seek portability now have real options. Users who don't will continue to experience every model release as a small bereavement.

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus

Methodology and sources

Events were selected on three criteria: (1) public announcement or contemporaneous coverage, (2) material impact on what users' assistants remembered or how they behaved, (3) reproducibility — the change could be observed by multiple independent users at the time. Internal A/B tests and silent backend rollouts are excluded unless they were later confirmed by the provider.

Primary sources: provider release notes, deprecation announcements, official blog posts, and contemporary coverage in The Verge, Ars Technica, TechCrunch, and the AI subreddits. Full source list available on request to hello@konshus.ai.

This report is updated quarterly. Last updated: June 2026.

Related

Frequently Asked Questions

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus