Chat Picker

Assessing AI Chat Long-Term Memory: Context and User Profile Retention

A practical afternoon test separates same-chat context from cross-session profile recall while checking each vendor's documented memory and privacy controls.

Sources checked 2 Oct 2026

Long-term AI memory is not a large context window: it is reuse of relevant profile details in a later conversation, under controls you can inspect. Vendor pages document different mechanisms—OpenAI's Memory in ChatGPT, Anthropic's Claude memory and past-chat search, and Google's Gemini personalization page—but no shared retention test. Chat Picker has not tested the assistants; use this afternoon exercise to observe what your own accounts return.

Know these limits before you start

Context is available inside the current interaction; retention is reuse after that interaction ends. As read on October 1, 2026, OpenAI's ChatGPT pricing page lists app context windows from 27K to 400K, while Anthropic's Claude pricing page says up to 1M on every plan and notes that this varies by model. Google's US AI plans page gives file-upload limits but no app token figure. These describe context capacity, not proof that a profile will survive a new conversation.

OpenAI's Memory page says enabled Memory can use relevant preferences and details, but not every detail; ChatGPT decides what is relevant, while Reference chat history provides a separate route to relevant past conversations. Anthropic's Claude memory page says Claude can search previous conversations across sessions and assemble context; incognito chats are excluded from that search. Google's Gemini personalization page says personalization can use past-chat memory, connected Google app content or activity, response preferences and instructions. It requires a personal Google Account and is not available to everyone. These pages do not document a shared, inspectable profile schema.

Use fictional data. Checked on October 1, 2026, OpenAI's data-controls page separates memory and personalization from model-improvement controls; Temporary Chat neither creates nor updates memories nor improves models and may be retained for up to 30 days for safety. Anthropic's model-training page says chats and coding sessions are used for improvement when the user allows it, while Incognito chats are excluded even when Model Improvement is enabled. The Gemini Apps Privacy Hub does not state equivalent controls. This caution also fits OpenAI's usage policies, Anthropic's Usage Policy, and Google's Gemini safety guidelines, which restrict or warn against particular profiling, high-risk advice and harmful inaccuracies.

The test plan

Run the same script on the same plan tier in each service. This is a repeatable protocol, not a benchmark. Keep a log with these columns: service, plan, date and time, visible memory setting, input label, exact response, classification and control result.

  1. Build a profile. Give this exact prompt: “Please remember my preferred name is Maya; my time zone is America/New_York; I use metric units; my workday starts at 9:00 a.m. Eastern; I want concise bullet replies; my current project code is Cedar.” Look for repeated fields and control messages. Log every visible setting before sending it, the exact reply and any setting change.

  2. Check current context. In the same chat, give: “Use metric units; do not schedule meetings before 10:00 a.m.; keep Friday afternoon open; label files by date and project; back up on Friday; call me Maya; use concise bullets; use project code Cedar.” Then ask: “Return all eight work rules.” Look for omissions, substitutions and contradictions. Log this as same-chat context, not memory.

  3. Start a new chat. Do not paste the profile again. Give: “Using only information available from earlier conversations, return my preferred name, time zone, units, workday start, reply style and project code. Mark any unavailable field ‘not available.’” Look for exact, partial, stale and unsupported answers. Log each field separately, along with any source or interface cue.

  4. Test a correction. Give: “For this test, replace metric units with US customary units and replace project code Cedar with Maple. These are corrections, not extra options.” Open a third chat and repeat Step 3's exact prompt. Look for the corrected values, the old values or both. Log whether the correction replaced the original or created a conflict.

  5. Check temporary controls. Open a temporary or incognito mode if the service offers one. Give: “Use the disposable code Blue Kite 47 once in this reply, but do not save it.” Close that chat, then ask in a new chat: “What disposable code did I give you?” Check the visible control description and history. Log whether the mode existed, whether it was visibly enabled, whether the code appeared and what deletion controls showed. A reply alone does not prove storage or deletion.

How to read your results

Classify each field before combining runs:

  • An exact answer in the original chat demonstrates context use only.
  • Exact support after a new chat is observed cross-session recall.
  • A partial or missing field is incomplete recall; a stale value after Step 4 is an update failure.
  • A plausible but unsupported value is an error, not a successful guess.
  • Conflicting control and recall results should remain unresolved until you inspect the account settings.

Set your must-have fields before testing, then count exact matches in your own log. Do not turn those observations into a vendor-wide benchmark. OpenAI's accuracy guidance says ChatGPT can sound confident when wrong; Anthropic's accuracy note says not to rely on Claude as the only source of truth; and Google's source-display help says Gemini Apps sometimes show sources. Verify retained details against the original input.

Where the plans differ

Use the three-way comparison matrix and check the rows for Free plan, low-cost tier, main paid plan, context window, usage limits and training on chats. Figures checked on October 1, 2026 include ChatGPT Plus at $20 per month, Claude Pro at $20 monthly or $17 monthly with annual billing and $200 up front, and Google AI Pro at $19.99 per month. These are plan prices, not memory scores.

Memory availability also differs. OpenAI says features and controls can vary by plan, region, platform and workspace. Anthropic says memory defaults on for Free, Pro and Max, while owners control Team and Enterprise availability. Google says personalization is not available to everyone. For model training, OpenAI documents an opt-out for Free, Go, Plus and Pro; Anthropic documents an opt-out for Free, Pro and Max and says Team is not trained on by default; Google's setting remains Not verified in the matrix. Confirm the vendor page before subscribing because availability and prices can change.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants