--- title: "Test your context window versus persistent AI memory." description: "Run same-session and new-session recall tests with harmless facts. Check sources, corrections, and export paths before relying on memory." canonical: "https://scalewithsearch.com/articles/context-window-vs-persistent-ai-memory" date: "2026-08-27" modified: "2026-09-25" --- ## Site navigation - [Scale With Search](https://scalewithsearch.com/) - Real estate - Real estate - [Real estate](https://scalewithsearch.com/for/real-estate) - Work - Start here - [Send your brief](https://scalewithsearch.com/work#send-your-brief) - [Prepare your six-question brief](https://scalewithsearch.com/work#prepare-your-six-question-brief) - Build - [Site build, content library with SEO, signal desk](https://scalewithsearch.com/work) - For your business - Trades and home services - [Auto body and collision shops](https://scalewithsearch.com/for/auto-body-and-collision-shops) - [Foundation and home repair contractors](https://scalewithsearch.com/for/foundation-and-home-repair) - [Garage door and fencing contractors](https://scalewithsearch.com/for/garage-door-and-fencing-contractors) - [HVAC contractors](https://scalewithsearch.com/for/hvac-contractors) - [Janitorial and commercial cleaning companies](https://scalewithsearch.com/for/janitorial-and-commercial-cleaning) - [Locksmiths](https://scalewithsearch.com/for/locksmiths) - [Moving companies](https://scalewithsearch.com/for/moving-companies) - [Pest control companies](https://scalewithsearch.com/for/pest-control-companies) - [Plumbing and electrical contractors](https://scalewithsearch.com/for/plumbing-and-electrical-contractors) - [Restoration and water or fire damage companies](https://scalewithsearch.com/for/restoration-and-water-fire-damage) - [Roofing companies](https://scalewithsearch.com/for/roofing-companies) - [Towing companies](https://scalewithsearch.com/for/towing-companies) - [Tree services and landscaping companies](https://scalewithsearch.com/for/tree-services-and-landscaping) - [Solar installers](https://scalewithsearch.com/for/solar-installation) - [General contractors](https://scalewithsearch.com/for/general-contractors-and-construction) - [Paving, concrete, and flooring contractors](https://scalewithsearch.com/for/paving) - Practices and professional services - [Bookkeeping and tax practices](https://scalewithsearch.com/for/bookkeeping-and-tax-practices) - [Dental practices](https://scalewithsearch.com/for/dental-practices) - [Family and criminal defense law firms](https://scalewithsearch.com/for/family-and-criminal-defense-law-firms) - [Med spas and aesthetics practices](https://scalewithsearch.com/for/med-spas-and-aesthetics) - [Personal injury law firms](https://scalewithsearch.com/for/personal-injury-law-firms) - [Veterinary clinics](https://scalewithsearch.com/for/veterinary-clinics) - [Gyms and fitness studios](https://scalewithsearch.com/for/fitness) - [Therapy and outpatient health practices](https://scalewithsearch.com/for/therapy-and-outpatient-health) - [Medical billing companies](https://scalewithsearch.com/for/medical-billing) - [Insurance agencies](https://scalewithsearch.com/for/insurance-agencies) - [Financial advisors](https://scalewithsearch.com/for/financial-advisors) - [Property management companies](https://scalewithsearch.com/for/property-management) - [Recruiting and staffing agencies](https://scalewithsearch.com/for/recruiting-and-staffing) - [Architects and interior designers](https://scalewithsearch.com/for/architects-and-interior-designers) - [Logistics and supply chain companies](https://scalewithsearch.com/for/logistics-and-supply-chain) - Agencies, MSPs, and manufacturing - [IT and managed service providers](https://scalewithsearch.com/for/it-and-managed-service-providers) - [Machine shops and precision manufacturers](https://scalewithsearch.com/for/machine-shops-and-precision-manufacturing) - [Marketing agencies and freelancers](https://scalewithsearch.com/for/marketing-agencies-and-freelancers) - [SEO agencies and consultants](https://scalewithsearch.com/for/seo-agencies-and-consultants) - [Small manufacturers and fabricators](https://scalewithsearch.com/for/small-manufacturers-and-fabricators) - Restaurants, shops, studios, and nonprofits - [Restaurants and hospitality businesses](https://scalewithsearch.com/for/restaurants-and-hospitality) - [Retail stores and ecommerce sellers](https://scalewithsearch.com/for/retail-and-ecommerce) - [Photographers, event planners, and travel agents](https://scalewithsearch.com/for/photographers) - [Churches and nonprofits](https://scalewithsearch.com/for/churches-and-nonprofits) - [All industries](https://scalewithsearch.com/for/) - Learn - For your office - [Office job guides](https://scalewithsearch.com/guides/) - [Browser calculators](https://scalewithsearch.com/tools/) - Start here - [How it works](https://scalewithsearch.com/how-it-works) - [Free Starter Kit](https://scalewithsearch.com/kit/business-memory-starter-kit.zip) - [Synthetic specimen](https://scalewithsearch.com/specimen/working-session-specimen.zip) - Guides - [The Complete Guide to Business Memory for AI Agents](https://scalewithsearch.com/articles/business-memory-for-ai-agents-guide) - [The Complete Small-Business Guide to AI Agent Governance](https://scalewithsearch.com/articles/ai-agent-governance-guide-small-business) - [The Complete Guide to Leaving Vendor AI Memory](https://scalewithsearch.com/articles/leaving-vendor-ai-memory-guide) - Articles by cluster - [Business memory](https://scalewithsearch.com/articles/business-memory-for-ai-agents-guide) - [Agent governance](https://scalewithsearch.com/articles/ai-agent-governance-guide-small-business) - [Migration and ownership](https://scalewithsearch.com/articles/leaving-vendor-ai-memory-guide) - For machines - [llms.txt](https://scalewithsearch.com/llms.txt) - [llms-full.txt](https://scalewithsearch.com/llms-full.txt) - [Machine view](https://scalewithsearch.com/?view=machine) - Company - Evidence - [Proof](https://scalewithsearch.com/proof) - Company - [About](https://scalewithsearch.com/about) # Context Window vs Persistent AI Memory: What Survives the Next Session. A buyer selects the model with the largest advertised context window. The team loads a long project packet, completes a planning session, and closes the chat. The next session does not know which options were rejected or which project state was approved. The model handled a large working set. The system did not preserve business state. A context window and persistent memory solve different problems. Context is the material available for the current response. Persistent memory makes selected state available after the current request or session. Chat history, retrieved documents, model training, and owned business records add more layers. Test them separately before buying on a token number. ## Treat context as the current working set If information disappears within the current thread, follow the [same-chat checkpoint procedure](/articles/why-chatgpt-forgets-same-chat) before changing cross-chat memory. Anthropic defines the context window as all text a model can reference while generating a response, including system instructions, messages, documents, tool definitions, tool results, and output. It calls this a form of working memory and warns that more context does not automatically produce better recall. [Anthropic: Context windows](https://platform.claude.com/docs/en/build-with-claude/context-windows) The exact capacity depends on model and product. Capacity can also be consumed by material users do not see as ordinary prose, including tools and prior output. Interfaces may compact or manage old content. Treat the context window as a bounded workspace. It is useful for: - reading a source packet; - comparing documents; - maintaining active reasoning state; - using current tool results; - producing one bounded output. It does not, by itself, promise that the next session will load the accepted decisions. Closing a chat, changing products, changing projects, or starting a fresh API request can create a new working set. Large windows also do not establish source authority. A stale policy and a current policy can both fit. The task still needs a precedence rule. ## See what happens when the working set overflows Capacity is measured in tokens, not words. Short common words are often one token, and longer words split into several. Code, technical terms, proper nouns, and non-English text usually need more tokens per word. A window therefore holds fewer words than its token count, and fewer still for code. Capacity fills faster than the interface suggests. One detailed exchange about a complex project can add thousands of tokens. A few uploaded documents can consume tens of thousands before the real work starts. A short question such as "What is a context window in AI?" is about nine tokens, depending on the tokenizer. In January 2026, Claude Opus 4.5 had a 200,000-token window. A 50-page document can take about 35,000 tokens of that before the first question. A medium-sized code repository can fill the whole window, and the model needs room to work as well as room to read. A context file of 10,000 tokens leaves about 190,000 tokens for the task itself. Window sizes have grown fast. A few years before 2026, 4,000 tokens was a common size. By January 2026, 200,000 tokens was common, and some models and betas offered a million. Check the current figure for the model you use. The limit exists because length costs compute. In the standard attention mechanism, the work grows with the square of the input length. A 200,000-token window needs about four times the attention computation of a 100,000-token window, not twice. The memory that holds the model's working state also grows with length. Longer windows therefore respond more slowly and cost more to run. The size of a window is an engineering trade between capacity, speed, and cost. The working-memory comparison is useful up to a point. A person holds only a small number of items in mind at once; George Miller's 1956 paper put the figure at about seven. A model's limit is measured in tokens instead of items, and everything counts against it: the prompt, earlier replies, corrections, follow-ups, and every shared file. When a conversation passes the limit, the product does not always warn. It may drop or compact the oldest content. Some systems cut from the middle instead: they keep the first prompt and the latest exchanges and drop what lies between. In both cases, a detail from half an hour earlier can vanish. The model behaves as if you never said it, because that text is no longer in its window. The earlier messages still appear on screen, and the user can scroll up and read them. The model may no longer receive them. Visible chat history is for the reader. The context window is what the model processes. A larger window does not remove three limits. Position is the first. Liu and colleagues found that performance is often highest when relevant information sits at the start or end of the input. It degrades when the model must use information in the middle of a long context. [Liu et al.: Lost in the Middle](https://arxiv.org/abs/2307.03172) A small packet of relevant sources often beats a large window filled with everything. Cost and speed are the second. Every token in the window is processed for every response. API users pay per input token, and larger inputs take longer. Session scope is the third. Even a million-token window covers only the current conversation. A new chat starts a new working set. Inside one session, four habits keep the working set useful: 1. Put critical context where the product keeps it, such as project instructions or a loaded context file, not only in an early chat message. 2. In a long thread, restate the key facts near the request, or write a checkpoint and continue in a new session from it. 3. Keep one topic per session, because tangents consume capacity. 4. Use headings and lists, so the model can locate each fact. Across sessions, the fix sits outside the window: records that load at the start of each session. Anthropic's Claude Code documentation states that each session begins with a fresh context window, and `CLAUDE.md` files carry instructions across sessions. [Claude Code: How Claude remembers your project](https://code.claude.com/docs/en/memory) The file stays on your disk and can sit under Git for change history. The text a session reads still goes to Anthropic for processing, and retention and training terms depend on the plan. The window becomes a view into a larger record, not the whole record. ## Define what persistent memory must preserve Persistent memory carries selected state across a defined boundary. The boundary may be requests, chats, sessions, projects, devices, or accounts. A vendor saying "memory" is not enough. Ask which boundary, which fields, which storage, which retrieval rule, and which correction path. Persistent state can include: - user preferences; - summaries of prior conversations; - project facts; - saved memories; - retrieved files; - application-managed records; - agent checkpoints. OpenAI's API provides Conversations objects to store and retrieve conversation state across Responses API calls. That is an application mechanism. The application still decides what belongs in the conversation and how long related state is retained. [OpenAI: Conversations API](https://platform.openai.com/docs/api-reference/conversations) Consumer product memory is another mechanism. OpenAI's ChatGPT Memory FAQ distinguishes saved memories from referenced chat history. [OpenAI: Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq) Saved memories are short summaries. They can hold that you prefer short emails, but not your full processes, client details, or decision rules. The business needs a stronger definition: current sources, accepted decisions, corrections, owners, approval evidence, and stopping rules survive in files or systems the business controls. The [business memory explainer](/articles/business-memory-for-ai-agents) defines that maintained operating record. ## Separate history, retrieval, and authority Chat history stores prior conversation text. Search can find it. That does not make every recovered statement current. Retrieval augmented generation finds selected material from a corpus and places it into current context. That does not decide whether the retrieved item is an archive, draft, correction, or accepted source. Model training changes model parameters through a training process. A fact seen during training is not a customer-controlled memory entry with an ordinary correction path. Prompt caching can reduce repeated input cost or latency. Anthropic's context documentation states that cached prompt prefixes still occupy the context window. Caching does not turn the input into a maintained business record. Use one term for each layer: - context window: current working set; - product memory: provider-managed cross-session state; - chat history: stored conversations; - retrieval: method for selecting source material; - authoritative record: maintained source that governs the field; - checkpoint: accepted task state used to restart work. The [RAG versus business memory guide](/articles/rag-vs-business-memory) explains why good retrieval still needs authority, correction, and expiry. ## Inspect `context-vs-memory-test.csv` Create a portable fixture with harmless facts: ```csv case_id,fact,source,in_window_recall,new_session_recall,correction_path,export_path,owner CVM-01,"Summary uses three bullets",preferences.md,pending,pending,edit preferences.md,exports/preferences.md,ROLE-USER CVM-02,"Project owner is Casey",projects/test/context.md,pending,pending,edit context.md,exports/project-test/,ROLE-PROJECT CVM-03,"Option Red was rejected",decisions/DEC-TEST-07.md,pending,pending,new decision plus supersession,exports/decisions/,ROLE-DECISION ``` Add model, product, account, project, context size, memory settings, retrieval settings, and time. Test each fact in the current session and after the target boundary. Do not expect every fact to persist. The test should state which layer is responsible. A preference may use product memory. A rejected commercial option should come from the decision file. An action approval should never be inferred from either. ## Run one same-session and one new-session test Use this test list. Keep the control account separate from the seeded fixture so prior history cannot supply the answer. 1. Run the same questions with no product memory and no retrieved files. Save the baseline result. 2. Start a clean session with the fixture and its sources. 3. Ask for each fact and require a source citation. 4. Add a long but harmless document between the fact and the recall prompt. 5. Repeat the prompt inside the same session. 6. Record omissions, wrong sources, and unsupported claims. 7. Close the session. 8. Start a new session under the same account and project. 9. Ask for each fact without supplying sources. 10. Record what product memory or history supplies. 11. Supply the owned source packet and repeat. 12. Correct one fact in the owned source. 13. Start another new session and verify the correction. 14. Replace the model or vendor and repeat from the same packet. 15. Confirm that no test result grants an external action. Pass conditions differ by layer. Same-session context passes when current sources are applied. Product memory passes only for the permitted fields. Business continuity passes when a clean replacement session reconstructs accepted state from owned records. The [long-document tool guide](/articles/best-ai-long-document-analysis-business) helps evaluate working-set performance. It should not be used as a proxy for persistence. ## Buy for the boundary, not token count Write the job before comparing products. Name the source size, session duration, boundary that state must cross, correction speed, export format, identity isolation, and consequence of stale recall. Ask vendors: - Does memory cross chats, projects, devices, and team members? - Can administrators scope or disable it? - Can users view, edit, export, and delete entries? - Does every entry retain provenance? - How are conflicts and superseded facts handled? - Can an owned source override recalled state? - What happens when the subscription ends? - Can another model reconstruct the result? A large window may be the right purchase for one-session document analysis. A smaller window with controlled retrieval and owned checkpoints may be better for recurring work. Many businesses need both. Do not buy persistence without testing deletion and correction. A system that remembers but cannot forget or show its source creates operating risk. ## Skip the persistent layer when the work is one-off A persistent layer is overhead in some cases. Casual questions, such as recipe or travel ideas, need little business context. One-off questions with nothing to repeat lose nothing when the model forgets. A single simple job, such as email drafts, may fit inside custom instructions or an existing Project. If your processes are not yet written down, write them first. A memory file depends on that work. Some people also prefer a blank slate for creative work, and that choice is valid. ## Approval and stopping boundary The evaluation may run synthetic same-session and new-session tests, inspect memory settings, record context capacity, and export test artifacts. It stops before uploading protected data, enabling organization-wide memory, changing retention, deleting business records, or acting on a recalled approval. The data owner approves persistence. The project owner approves project sources. The action owner approves external effects. If the tested boundary, source authority, account identity, or correction path is unclear, mark the product result unresolved and stop. ## Sources - [Anthropic: Context windows](https://platform.claude.com/docs/en/build-with-claude/context-windows) - [OpenAI: Conversations API](https://platform.openai.com/docs/api-reference/conversations) - [OpenAI: Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq) For the next step, read the [AI agent acceptance checklist for buyers](/articles/ai-agent-acceptance-checklist); [Full scope and terms](/work) covers the scoped build. ## Related: Owned Memory - [How to Reset AI Context Without Deleting the Business Record](/articles/reset-ai-context-without-deleting-memory) - [Why ChatGPT Forgets Things in the Same Chat and How to Preserve the Working State](/articles/why-chatgpt-forgets-same-chat) - [Compare eight AI context limit workarounds and keep the ones that last](/articles/ai-context-limit-workarounds) ---- ```text .|########||. .|########||. .|########||. |##||. .||##|. |##||. .||##|. |##||. .||##|. |#|. .|#|. |#|. .|#|. |#|. .|#|. |#| |#| |#| |#| |#| |#| .#| |#. .#| |#. .#| |#. .#. .#| .#. .#| .#. .#| |#. .#| |#. .#| |#. .#| |# ...... #| |# ...... #| |# ...... #| .# ||#########| #| .# ||#########| #| .# ||#########| #| .#. |######||######|. .#. .#. |######||######|. .#. .#. |######||######|. .#. #. .##|###|##|#|######| .# #. .##|###|##|#|######| .# #. .##|###|##|#|######| .# || |##|#||||||||||||#||#| || || |##|#||||||||||||#||#| || || |##|#||||||||||||#||#| || # |#||||||||||||||||||||#| #. # |#||||||||||||||||||||#| #. # |#||||||||||||||||||||#| #. || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || # .#||||||||||||||||||||||||#| # # .#||||||||||||||||||||||||#| # # .#||||||||||||||||||||||||#| # || ....|||#||||||##|#|||#|#||##||||....|. || || ....|||#||||||##|#|||#|#||##||||....|. || || ....|||#||||||##|#|||#|#||##||||....|. || #. . ....|# ....#||| ||| .#||| ....#. .# #. . ....|# ....#||| ||| .#||| ....#. .# #. . ....|# ....#||| ||| .#||| ....#. .# # . ||||#| .#####|| . .#| .#||| ||||# #. # . ||||#| .#####|| . .#| .#||| ||||# #. # . ||||#| .#####|| . .#| .#||| ||||# #. .| ||||| #. |#|||||. || #. |#||. .||||| || .| ||||| #. |#|||||. || #. |#||. .||||| || .| ||||| #. |#|||||. || #. |#||. .||||| || || |.... #. ....|#. |. ...|. ....|| || || |.... #. ....|#. |. ...|. ....|| || || |.... #. ....|#. |. ...|. ....|| || #. .||||||##||||||##||####|||||||#||||||#| .# #. .||||||##||||||##||####|||||||#||||||#| .# #. .||||||##||||||##||####|||||||#||||||#| .# # .||####|###########################|. # # .||####|###########################|. # # .||####|###########################|. # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#||||||############|######|||||#| # # .#||||||############|######|||||#| # # .#||||||############|######|||||#| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# || | ...|| ...##| | #| .|. |. ##### .. .| || || | ...|| ...##| | #| .|. |. ##### .. .| || || | ...|| ...##| | #| .|. |. ##### .. .| || || ||||| || ||||#| . |. |. |#. ||||| |#| || || || ||||| || ||||#| . |. |. |#. ||||| |#| || || || ||||| || ||||#| . |. |. |#. ||||| |#| || || .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. #.|||||||#############################| ||| .# #.|||||||#############################| ||| .# #.|||||||#############################| ||| .# || ##||#||#||||||#||#||#||#||##. || || ##||#||#||||||#||#||#||#||##. || || ##||#||#||||||#||#||#||#||##. || # .#|||||||||#||#|||||||||||#| # # .#|||||||||#||#|||||||||||#| # # .#|||||||||#||#|||||||||||#| # || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || .# |#||||||||||||||||||||#| #. .# |#||||||||||||||||||||#| #. .# |#||||||||||||||||||||#| #. || |#||||||||||||||||||#| || || |#||||||||||||||||||#| || || |#||||||||||||||||||#| || #. .######|#|#########| .# #. .######|#|#########| .# #. .######|#|#########| .# .#. |#####||||#####| .#. .#. |#####||||#####| .#. .#. |#####||||#####| .#. |# ||########|| #. |# ||########|| #. |# ||########|| #. |# ...... #| |# ...... #| |# ...... #| |#. .#| |#. .#| |#. .#| |#. .#. |#. .#. |#. .#. .#| |#. .#| |#. .#| |#. |#| |#| |#| |#| |#| |#| .|#|. .|#|. .|#|. .|#|. .|#|. .|#|. |##||. .||##| |##||. .||##| |##||. .||##| .||########||. .||########||. .||########||. Scale With Search 2026 [scalewithsearch.com](https://scalewithsearch.com) ```