Claude, ChatGPT, Gemini Memory Systems Compared: Entries vs States vs Data Sources

Overview

Over a three‑month period, Anthropic, OpenAI, and Google each overhauled their AI memory systems. Though the changes appear different, they all answer the same core question: what is the smallest actionable unit of memory? Claude chose discrete entries , ChatGPT chose a synthesized state , and Gemini chose connected data sources . This choice directly shapes how much users can inspect, correct, or delete what the model remembers.

Claude: From Daily Summaries to Discrete Entries

Granularity shift (July 10, 2026)

Before July 10, Claude carried a single daily summary — a compiled text block rebuilt each day. After the update, memory became a set of independent, category‑organized entries that are read and updated in real time during conversations.

Three consequences

Update timing: from post‑session compilation to in‑conversation writes. A preference expressed at 10 AM is usable by 2 PM.

Categorized entries: the memory panel reads like a checklist, not a monologue.

Per‑entry control: each entry can be viewed, edited, or deleted individually in Settings > Memory without wiping the entire memory.

The article illustrates the failure mode of summaries: an outdated fact (e.g., a job‑search inference from March persisting in July) forces a choice between resetting the whole memory or tolerating the error. With entries, the stale item is removed while the rest — writing style, tech stack, project context — survives. The analogy: password managers store discrete records, not a single text file; granularity is the line between a maintainable store and one you periodically nuke.

Reflect and unified memory (July 9 & August 25)

On July 9, Anthropic launched Reflect , a monthly review of topics, active days, peak hours, and task habits (beta for Free/Pro/Max on web/desktop, only when memory is enabled). Two days later, entries replaced summaries. On August 25, memory was unified across Chat and cloud Cowork: entries are saved by topic in real time, visible under Topics, and shared only with cloud Cowork (local Cowork, incognito chats, and Claude Code use separate stores). Sensitive topics (health, race, religion, politics, gender) are excluded by default; a toggle enables them prospectively. Hard‑blocked data (government IDs, financial accounts, criminal/immigration records) is never stored. Users can pause (retain but stop read/write) or reset (permanently delete all entries). A notable side effect: deleting a conversation no longer deletes its derived memory entries — Anthropic’s help docs state this explicitly, yet the UI lacks change history, source‑conversation backlinks, or a separate off switch for Reflect.

ChatGPT: Dual System and Dreaming

Two‑layer architecture

Saved Memories (April 2024): explicit “remember this” notes written only during conversation, triggered by strong prompts. OpenAI admits it feels like talking to someone who remembers a few notes but forgets everything else; memories grow stale and become incorrect.

Dreaming (April 2025 → V3 on June 4, 2026): a background process that reads historical chats, merges duplicates and contradictions, and writes an integrated memory state. V3 runs in idle compute windows, reducing per‑user compute to ~1/5 of the previous cost, enabling free‑tier coverage.

Key capability: auto‑rewrite

A memory like “you’re traveling to Singapore in July” silently becomes “you traveled to Singapore in July 2026” after the trip, without user intervention. OpenAI’s self‑reported metrics (methodology and datasets not public): factual recall 41.5 % → 82.8 %; preference adherence 71.3 %; time‑sensitive accuracy 9.4 % → 75.1 % (the largest single improvement). This solves real pain points — e.g., “preparing a spring wedding” or “toddler at home” no longer persist as true after the event.

Trade‑off: loss of auditability

The memory summary page shows highlights and allows additions/updates, but it is not the full database; low‑relevance or sensitive details are omitted. Deleting a memory entry is not durable because the background process can re‑synthesize it from surviving source conversations. OpenAI’s guidance: to fully remove something, delete both the memory and the source chats.

Project‑only memory (August 14, 2026)

Project‑only memory (introduced August 2025) can now be toggled post‑creation via Project settings > Memory. It isolates context: the project uses only its own chats, never updates account‑level memory, and never pulls from outside. Limitations: only for non‑shared projects (shared projects are permanently project‑only); ChatGPT Work unavailable in project‑only projects; no global default — each project must be configured individually. On non‑Enterprise plans, both Reference saved memories and Reference chat history must be on; turning off the former disables the latter, and turning off the latter purges historically learned info within 30 days. Temporary chats bypass history, memory, and training entirely.

Gemini: Data‑Source Approach

Gemini’s Personal Intelligence (launched January 2026 in the US, later global) connects to Gmail, Calendar, Docs, Drive, Keep, Tasks, Photos, YouTube, YouTube Music, Spotify, Maps, Flights/Hotels, SMS/WhatsApp, Contacts, Google Home, GitHub, Dropbox, OpenTable, etc. Default off; Google pledges not to train on Gmail/Photos.

Four practical pitfalls

Fragmented tiers: context windows split by plan — no plan 32K tokens, AI Plus 128K, AI Pro/Ultra 1M. “Million‑token context” is misleading; availability depends on product, model, plan, and feature. Explicit memory works on free tier, but automatic learning from history requires a paid subscription.

Regional exclusions: UK, Nigeria, South Korea, Switzerland, and the entire European Economic Area are excluded; paid tiers roll out first, free users wait weeks. Companion features like Computer History are similarly absent in EEA/UK/Switzerland.

Premature forgetting: June 2026 paid users reported active‑session amnesia. Despite advertised 1M‑token windows, dynamic conversation windows compress to ~16K tokens — roughly 25‑30 messages — causing loss of early parameters, code blocks, and structural constraints. The same 1M window works correctly in Google AI Studio, highlighting a consumer‑app regression.

Ecosystem lock‑in: Gemini memory stays inside Gemini. Context accumulated there does not transfer to Claude, ChatGPT, or other models; switching costs equal the accumulated context, not the subscription fee.

Structural privacy note: Gemini memory resides alongside email, search history, location, YouTube history, photos, calendar, and docs in the same Google account. Consumer conversations may be used for model improvement unless Gemini Apps Activity is disabled; human reviewers may read/annotate conversations. Workspace/Enterprise plans use a separate data‑processing regime. Advantage: user‑configurable retention periods (finest granularity of the three) and comprehensive Takeout export, though export formats are Google‑centric.

User‑Perspective Ranking

Claude — most reliable. Only system with a per‑entry auditable list. Granularity enables single‑error fixes; real‑time writes make morning preferences available afternoon; clear scope boundaries (account, cloud Cowork, local Cowork, project, incognito, Claude Code); sensitive‑topic defaults and hard blocks keep high‑risk data out. Cost: user must actively review Settings > Memory and remember “delete chat ≠ delete memory”.

ChatGPT — most understandable. Summary page gives a clear view; add/update/topic‑timing controls in one UI. Dreaming’s auto‑rewrite solves memory rot (time‑sensitive accuracy 9.4 % → 75.1 %). Cost: no per‑entry audit; memory becomes a state, not a list; deletions are non‑durable because source chats can regenerate them; rollout is phased (US Plus/Pro first).

Gemini — least stable. Richest data sources, highest theoretical ceiling, but fragmented tiers, regional gaps, ~16K dynamic window causing 25‑30‑message amnesia, and full ecosystem lock‑in. Configurable retention is a genuine strength but cannot compensate for “forgets mid‑conversation”.

In one sentence: Claude lets you manage, ChatGPT lets you understand, Gemini lets you use but not necessarily remember.

Actionable Recommendations

Read your memory panel this week. Open Settings > Memory / Topics (Claude), Memory Summary (ChatGPT), or Connect apps (Gemini). An unread memory is not a feature — it’s a risk exposure.

Isolate by sensitivity, not by topic. Create separate projects for client work, compensation/HR, health; enable project‑only memory. General writing and learning stay in the default scope. This beats any privacy toggle.

Treat deletion as a two‑step operation. In ChatGPT: delete both the memory entry and the source chats. In Claude: delete the memory entry itself, not the conversation. In Gemini: disconnect the linked app and clear stored content in Gemini Apps Activity — disconnecting alone does not erase already‑saved data.

Memory is shifting from “nice‑to‑have” to “infrastructure”. Three months ago we compared summary quality; today we compare who lets you see what’s stored and who lets you make it forget.