LEED Agency's reporting problem was not a shortage of AI answers. It was the loss of context between answers. Each account carried its own Google Ads rules, assets, metrics, screenshots, formats and history. A useful system had to retrieve that context before analysis and retain what the work revealed afterward.
Evidence at publication: the runtime was verified live on 30 May 2026; recent logs showed work inside real reporting flows; James described the system as part of his daily operation.
The company
James runs LEED Agency, working across Google Ads, e-commerce reporting, client accounts and product ventures. The operational file base spans active reports, account knowledge, proposals, audits, Webflow material, scratch work and product context.
That breadth matters. A report for one client cannot safely inherit another client's assumptions. The value of the work lives in small, accumulated distinctions: which metrics matter, which rules are non-negotiable, where source assets live, how a report should read and what the last analysis discovered.
The constraint
Before the memory and migration work, TheoAI could behave like a fresh model even when the right knowledge existed elsewhere. Longer, multi-step Google Ads reports could time out. James then had to reconstruct the account, restate the rules and reassemble the reporting context before useful analysis could begin.
The bottleneck was not merely retrieval. It was continuity. Work completed in one session needed to improve the next one instead of disappearing into chat history.
Why existing tools were not enough
Files could store reports. Chat could answer questions. A browser could inspect sources. None of those surfaces alone knew which LEED client was active, which SOP governed the task or which prior finding belonged in memory.
The missing operating layer had to organize the workspace, retrieve account-specific knowledge, preserve James's tone and hard rules, allow enough time for deeper work, fail cleanly when a task stalled, and write useful new context back into the system. Without those controls, a powerful model remained an unreliable reporting assistant.
Transformation map
Before
Request -> current chat context -> manual file hunt -> partial analysis -> timeout or report -> context lost
After
Request -> identify client and reporting rule -> retrieve LEED memory and source files -> analyse -> James review -> report asset -> retain useful finding
The model is one component. The operating improvement comes from routing, memory, workspace structure, timeouts and a clear human review point.
One workflow, end to end
- Input: James requests a report or account analysis in an established Discord or Telegram channel.
- Context selection: Theo identifies the relevant LEED client, product or venture and retrieves its memory, SOPs and report format.
- Source inspection: The assistant works through the appropriate files and browser-accessible material instead of answering from the current message alone.
- Analysis: The longer task runs with an extended timeout and a finite idle watchdog, allowing depth without leaving a failed channel indefinitely stuck.
- Human gate: James reviews the analysis and retains authority over what enters a client-facing report. The source does not claim autonomous client delivery.
- Output: The report or supporting asset is saved into the organized client workspace.
- Record: A durable finding can be written back into memory, making later work start with more context.
What changed
- Built: Persistent LEED knowledge, organized client and venture workspaces, report-safe timeout behavior, memory search and voice support.
- Connected: Discord, Telegram, local speech-to-text, text-to-speech, browser control and the agency's structured knowledge base.
- Tested: Gateway health, model routing, active communication surfaces, browser service, report timeout behavior and workspace retrieval were inspected.
- Verified live: On 30 May 2026, the gateway LaunchAgent was running, the local health endpoint returned live, Discord and Telegram were enabled and the browser service was active.
- Actively used: Recent live logs showed Theo working in Google Ads/Calendly conversion analysis and report-asset handling. James's approved remarks describe daily use.
- Provider or client dependency remaining: No missing provider is presented as blocking the retained reporting workflow. Like every operational case, current state is reviewed quarterly.
- Financial result reported by client: None. This case is evidence of operational use and context continuity, not a revenue attribution claim.
Proof
- Live runtime: OpenClaw gateway, browser service, Discord and Telegram verified on 30 May 2026.
- Reporting control: Task timeout extended to 1,800 seconds, with finite idle watchdogs for clean failure.
- Knowledge depth: The retained workspace includes multiple LEED accounts, proposals, audits, product ventures and report assets.
- Use evidence: Recent logs showed ongoing reporting activity rather than an installed but idle system.
- Client corroboration: Two exact, permission-approved remarks are retained in the claims ledger.
Technology - revealed last
The system runs on James's Mac mini with OpenClaw, Discord, Telegram, Codex model routing, GPT-5.4 at the recorded verification point, local whisper.cpp speech-to-text, Edge TTS, ffmpeg, browser control, ocbrain, memory search, dreaming and launchd service management. The workspace is organized into clients, ventures, knowledge, memory, proposals and scratch areas so retrieval follows the agency's work rather than a generic chat taxonomy.
Client quote
“It’s changed my life literally.”
“I pretty much use it every single day, every second of the day.”
- James, LEED Agency. Exact wording and publication approval are recorded in the claims ledger; the quote source is the founder-approved website blueprint.
Next transformation
See how Plutify turned accounting data into a live, read-only client review surface. Both cases address reporting friction, but Plutify solves the client interface while LEED solves the agency's working memory.
Get your AI blueprint
If reporting quality depends on one person remembering every account rule, the first transformation is usually the context and evidence path - not report generation alone. Book a free AI audit to map it.
Source record
OPENCLAW-DEEP-CLIENT-CASE-STUDIES.pdf, page 13.RECAST_CASE_STUDIES_2026_UPDATED.pdf, page 7, documentRA-CASES-001, version 2.0.RECAST_WEBSITE_MASTER_BLUEPRINT.md, sections 6.5, 10.4, 11 and 58.