--- title: "How Corrections Become Part of the System" description: "Record the wrong behavior, approved rule, source, scope, and test so a correction outlives the chat and future agents apply it." canonical: "https://scalewithsearch.com/articles/corrections-become-system-memory" date: "2026-08-17" modified: "2026-10-03" --- ## 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) # How Corrections Become Part of the System. An agent uses the wrong offer name. You correct it in the chat. The next session uses the wrong name again because the correction never left the conversation. Feedback becomes system memory only when it changes a durable record, retrieval path, rule, or test. "Remember this" is not enough. The system needs to know what was wrong, what the approved rule is, where it applies, and how a future run proves compliance. ## Why a chat correction does not carry over A correction typed into a chat changes that chat. It reaches the next session only when something loads it. Product memory features save some corrections on the product's terms. OpenAI's ChatGPT Memory FAQ describes saved memories that the user can review and delete. The product still decides what to save and when to apply it. [OpenAI: Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq) The guide to [correcting a wrong ChatGPT memory](/articles/correct-wrong-chatgpt-memory) traces a bad saved value to its layer. Claude Projects keep the instructions and files that someone adds, and project knowledge changes only when someone edits it. A custom GPT changes behavior only when its configuration changes. The [Claude Projects and custom GPTs comparison](/articles/claude-projects-vs-custom-gpts-business-memory) shows the manual update path in each. Claude Code has an auto memory that writes notes from corrections and preferences into a local memory directory. It loads the first 200 lines of the `MEMORY.md` index, up to 25KB, at the start of each session (checked 2026.09.25). [Claude Code: How Claude remembers your project](https://code.claude.com/docs/en/memory) None of these mechanisms gives the business a reviewed rule with a source, a scope, and a test. The workflow on this page closes that gap. The failure pattern is common. In session 1, the agent writes dates as MM/DD/YYYY, and the owner corrects it in the chat. In session 2, the agent repeats the error, because the correction never left session 1. Then one line goes into the controlling file: ```text date_format:: YYYY.MM.DD ``` Session 2 reads the file and applies the rule. Session 50 still applies it, because the rule is written down, not buried in a transcript. Reusable corrections of this kind fall into a few classes: - format: date style, heading case, list style, client name style; - domain terms: "use COE, not close of escrow"; - workflow rules: "confirm before deleting any file"; - banned phrases: a list the writer checks before every draft. A written correction is also portable. When the business changes model or vendor, the file moves with it, and the replacement reads the same rule. A loaded rule file also steadies the output. Without it, each session guesses from the phrasing of the request, so the same question can get different answers. With the file loaded, the assumptions stay the same from run to run. The smallest working version is one `CLAUDE.md` file that Claude Code reads at the start of each session. Add each accepted rule there, and split the file when it grows. The file is the easy part. The real work is to write the business context clearly enough for a model to apply it. The rest of this page adds the review that keeps the file trustworthy. ## Capture the correction at the right level Some corrections are local: "Change this heading to sentence case." Some corrections are reusable: "All public pages use the approved offer names from the canonical offer file." Some corrections are decisions: "The owner approved a new delivery date for Client A." Do not store these as if they have equal scope. A local edit belongs in the current artifact. A reusable behavior belongs in `corrections.md` or a runbook. A business decision belongs in the authoritative decision or offer record, with the correction pointing to it. ## Record five parts A reusable correction needs: 1. the observed wrong behavior; 2. the approved rule; 3. the source of authority; 4. the scope where it applies; 5. a test for the next run. For example: ```text correction_id:: COR-027 example_type:: synthetic observed_in:: receipts/RUN-2026-0817-0091.json wrong_behavior:: omitted the current delivery exception approved_rule:: include active exceptions listed in account-context.md source:: accepted account decision DEC-014 scope:: weekly account brief for CLIENT-A controlling_object:: briefs/weekly-account-brief.md test_added:: tests/weekly-account-brief/active-exception.yaml expected_result:: current exception appears with source DEC-014 status:: proposed; acceptance and retest still required ``` This gives a future agent something operational. It can read the rule and check the artifact. The [correction log template](/articles/ai-correction-log-template) adds identifiers, affected outputs, and retest results when several workflows share a field. ## Correct the source, not every output If ten pages repeat an outdated price, the primary fix belongs in the canonical offer record. Then the downstream pages can be regenerated or reviewed against it. Editing ten outputs without correcting the source guarantees another recurrence. The correction record should point to the authoritative file. It should not become a competing copy of every fact. This is why provenance matters. A correction without a source can become another unsupported instruction. ## Use diffs for review Text files make proposed corrections visible. Git's diff tools show line-level and word-level changes, which helps a reviewer distinguish a changed rule from surrounding formatting. [Git diff documentation](https://git-scm.com/docs/git-diff) The review object is the exact change: ```diff - delivery_day:: Tuesday + delivery_day:: Thursday ``` The reviewer can accept, reject, or amend that line. The accepted file then becomes the source for later work. Version history also helps answer when a rule changed and which artifacts may need rechecking. Git's log documentation covers history inspection and file-specific traversal. [Git log documentation](https://git-scm.com/docs/git-log) ## Inspect `corrections.md` Choose one repeated failure. Add it to `corrections.md` with the five required parts. Then run a regression test: 1. Start a fresh agent session. 2. Provide the task brief, source record, and corrections file. 3. Ask for the same type of output that failed. 4. Check the named test. 5. Write the result to `receipts/correction-test.md`. Pass means the new output follows the approved rule and cites the governing file. Fail means the correction is unclear, not retrieved, or attached at the wrong level. Do not mark the problem fixed because one chat response improved. Mark it fixed when a cold run passes. ## Route the correction to the controlling object A correction should change the smallest authoritative object that can prevent recurrence. If the wrong recipient appeared because identity resolution failed, update the identity rule and its test. If an accepted format was ignored, update the task brief or approved example. If the system used an old business fact, update the source record. If the tool performed an unapproved action, update the capability gate and denied-action test. Writing every issue into one `corrections.md` file creates a second policy system. Use the file as an intake and traceability surface, then route stable fixes to the object that controls the behavior. For synthetic `COR-027` above, the brief is the controlling object and the active-exception fixture proves the change. Keep acceptance and retest evidence beside the entry. ## Distinguish a correction from a new request Not every revision means the system failed. "Use the approved account name" corrects a factual or policy error. "Try a shorter introduction on this draft" may be a local editorial request. "Add a new approval stage" changes the workflow. "Prepare a second report" expands scope. Classify the feedback before it enters durable memory: - defect correction: expected behavior was defined and missed; - source update: the business fact changed; - preference: a reviewer wants a different presentation; - workflow change: the contract, tool, or boundary changes; - scope request: a new result is being requested. Only the first two normally enter the current correction path without a new design decision. Workflow and scope changes need their own acceptance criteria, owner, and review. Some repeated failures are not memory problems. A rule file adds context, not subject knowledge. If the model does not know the subject, a correction file does not fix that gap. A vague prompt also produces vague output. "Make this better" gives the model little to act on. "Rewrite this email to be more direct and keep my usual sign-off" states the result. ## Retest at the same boundary that failed A corrected draft can hide an unfixed system. If a person manually edits the output, the artifact improves while the next run remains unchanged. Rerun from a fresh session with the same class of input. Load the accepted source and correction. Exercise the exact boundary that failed. If the failure involved a live capability, use a recorder or test account instead of repeating the external effect. The [run receipt](/articles/receipt-is-part-of-the-output) should state why the system passed: a changed controlling file, retrieval path, or adapter, or a human repair of the artifact. Only the first three can support a claim that the workflow itself improved. ## Prevent correction bloat An endless correction file becomes another archive. Review active entries. Merge duplicates. Move superseded rules to history. Promote stable rules into the main runbook or source record. Keep the correction file focused on behavior the system still needs to remember. Add fields such as `status:: active`, `superseded_by::`, and `last_tested::`. The goal is a maintained control surface, not a museum of every edit. Set a review owner and cadence in the runbook. During review, compare active corrections with the source record and current fixtures. Archive only entries whose replacement is named and tested. If a rule has no owner, no scope, or no recent test, flag it for a decision instead of applying it silently. ## Separate taste from authority Some feedback expresses a preference for one artifact. Other feedback changes business policy. Before promoting a correction, ask: - Should this apply next time? - Which jobs should inherit it? - Who may approve it? - Does an existing source already govern it? - What would prove compliance? If the answer is unclear, keep the change local until the owner decides. ## Approval and stopping boundary An agent may propose correction records and show the exact diff. It must not make a new price, contract term, legal position, permission, personnel rule, or public claim authoritative without owner approval. When feedback conflicts with an active canonical rule, the system stops. It presents both records and asks which one should govern. No silent overwrite. No policy by accident. ## Keep the correction with the work A proposal says one revision round. Your current terms say two. You fix the proposal, then find the old wording in another draft. The useful object is the decision behind that edit. Save what changed, where the decision came from, and which work it governs. Include what the correction does not cover. Website revision terms should not silently become the terms for every service you deliver. Give the next task the current record and ask it to identify the applicable rule before drafting. Check the cited source. Then check the actual output. A correct citation does not establish correct application. Compare each condition and exception with the source. A worksheet can cite the right decision while dropping a condition or inventing a threshold. If the source leaves something unresolved, record "not established" and a question for the owner. Do not fill the gap to complete a template. Preserve an old example and the corrected version. Use them when changing tools, revising instructions, or handing work to another person. Keep approval separate. A paragraph saying "ask before sending" is an instruction, not proof that a tool cannot send. ### Practice 1. Select one decision you have already made and one draft that still contradicts it. 2. Find the original decision and preserve its exact reference. 3. Complete the correction record below. Define the applicable work and the exceptions. 4. Retrieve that record using the question a future task would ask. 5. Revise the old draft, then compare the output with the source and exceptions. 6. Ask another person to explain the rule without your verbal help. Record any ambiguity they find. 7. Save the accepted revision and the remaining question. Keep any external action separate. Expected artifact: one correction record, one checked revision, and a clear next-use instruction. ## Copyable correction record ```markdown # Correction: [decision name] source:: [exact decision file and section] decided:: [date] owner:: [person who can change this decision] status:: [proposed or accepted] Current rule: [what the source establishes] Applies to: [specific work, audience, and context] Does not apply to: [exceptions and unrelated work] Previous wording: [the old claim, preserved as history] Reason: [source-supported reason, or not recorded] Affected materials: [exact drafts, pages, templates, or instructions] Retrieval question: [how the next worker would ask for this rule] Retrieved source: [actual result, not intended result] Condition check: [source conditions and exceptions retained; missing information labeled] Checked example: [old input and corrected output] Review: [reviewer, date, accepted changes, remaining questions] External action: [none, or separate authorization reference] Next use: [which task must consult this record] ``` ## A correction is complete when the next run changes Durable memory is visible in future behavior. Capture the failure. Correct the authoritative source. Add a scoped rule when needed. Review the diff. Run the regression test. Keep the receipt. ## Questions about How Corrections Become Part of the System ### Why does an AI assistant repeat a mistake after an evidence-based correction? Record the observed wrong behavior, approved rule, source of authority, scope, and a test for the next run. Then attach the correction to the controlling source or workflow object and rerun the same class of task from a fresh session. ### How do I turn a correction into a rule I can test in the next session? A chat acknowledgment does not prove that the system changed. Put a reusable correction in corrections.md or the authoritative source, record its scope and approval, and use a regression test to verify that future output follows it. ### How do I make ChatGPT apply my correction rule consistently across conversations? Separate a local editorial preference from a reusable behavior rule or a business decision. Promote only the durable rule, route it to the smallest authoritative object that can prevent recurrence, and retain a receipt showing whether the cold run passed. ## Save the visual summary Record five parts | Route the correction to the controlling object | Retest at the same boundary that failed | Prevent correction bloat [Download the PNG](/infographics/corrections-become-system-memory-1200x1500.png) ## Related: Owned Memory - [What Context Should an Agent Read Before It Starts?](/articles/what-context-should-an-agent-read) - [Measure the time you lose to AI context setup before you fix it](/articles/how-much-time-wasted-on-ai-context) - [Document your business processes so an AI can follow them without a briefing](/articles/how-to-document-business-processes-for-ai) ---- ```text .|########||. .|########||. .|########||. |##||. .||##|. |##||. .||##|. |##||. .||##|. |#|. .|#|. |#|. .|#|. |#|. .|#|. |#| |#| |#| |#| |#| |#| .#| |#. .#| |#. .#| |#. .#. .#| .#. .#| .#. .#| |#. .#| |#. .#| |#. .#| |# ...... #| |# ...... #| |# ...... #| .# ||#########| #| .# ||#########| #| .# ||#########| #| .#. |######||######|. .#. .#. |######||######|. .#. .#. |######||######|. .#. #. .##|###|##|#|######| .# #. .##|###|##|#|######| .# #. .##|###|##|#|######| .# || |##|#||||||||||||#||#| || || |##|#||||||||||||#||#| || || |##|#||||||||||||#||#| || # |#||||||||||||||||||||#| #. # |#||||||||||||||||||||#| #. # |#||||||||||||||||||||#| #. || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || # .#||||||||||||||||||||||||#| # # .#||||||||||||||||||||||||#| # # .#||||||||||||||||||||||||#| # || ....|||#||||||##|#|||#|#||##||||....|. || || ....|||#||||||##|#|||#|#||##||||....|. || || ....|||#||||||##|#|||#|#||##||||....|. || #. . ....|# ....#||| ||| .#||| ....#. .# #. . ....|# ....#||| ||| .#||| ....#. .# #. . ....|# ....#||| ||| .#||| ....#. .# # . ||||#| .#####|| . .#| .#||| ||||# #. # . ||||#| .#####|| . .#| .#||| ||||# #. # . ||||#| .#####|| . .#| .#||| ||||# #. .| ||||| #. |#|||||. || #. |#||. .||||| || .| ||||| #. |#|||||. || #. |#||. .||||| || .| ||||| #. |#|||||. || #. |#||. .||||| || || |.... #. ....|#. |. ...|. ....|| || || |.... #. ....|#. |. ...|. ....|| || || |.... #. ....|#. |. ...|. ....|| || #. .||||||##||||||##||####|||||||#||||||#| .# #. .||||||##||||||##||####|||||||#||||||#| .# #. .||||||##||||||##||####|||||||#||||||#| .# # .||####|###########################|. # # .||####|###########################|. # # .||####|###########################|. # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#||||||############|######|||||#| # # .#||||||############|######|||||#| # # .#||||||############|######|||||#| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# || | ...|| ...##| | #| .|. |. ##### .. .| || || | ...|| ...##| | #| .|. |. ##### .. .| || || | ...|| ...##| | #| .|. |. ##### .. .| || || ||||| || ||||#| . |. |. |#. ||||| |#| || || || ||||| || ||||#| . |. |. |#. ||||| |#| || || || ||||| || ||||#| . |. |. |#. ||||| |#| || || .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. #.|||||||#############################| ||| .# #.|||||||#############################| ||| .# #.|||||||#############################| ||| .# || ##||#||#||||||#||#||#||#||##. || || ##||#||#||||||#||#||#||#||##. || || ##||#||#||||||#||#||#||#||##. || # .#|||||||||#||#|||||||||||#| # # .#|||||||||#||#|||||||||||#| # # .#|||||||||#||#|||||||||||#| # || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || .# |#||||||||||||||||||||#| #. .# |#||||||||||||||||||||#| #. .# |#||||||||||||||||||||#| #. || |#||||||||||||||||||#| || || |#||||||||||||||||||#| || || |#||||||||||||||||||#| || #. .######|#|#########| .# #. .######|#|#########| .# #. .######|#|#########| .# .#. |#####||||#####| .#. .#. |#####||||#####| .#. .#. |#####||||#####| .#. |# ||########|| #. |# ||########|| #. |# ||########|| #. |# ...... #| |# ...... #| |# ...... #| |#. .#| |#. .#| |#. .#| |#. .#. |#. .#. |#. .#. .#| |#. .#| |#. .#| |#. |#| |#| |#| |#| |#| |#| .|#|. .|#|. .|#|. .|#|. .|#|. .|#|. |##||. .||##| |##||. .||##| |##||. .||##| .||########||. .||########||. .||########||. Scale With Search 2026 [scalewithsearch.com](https://scalewithsearch.com) ```