Keva Brain
Most AI support agents have amnesia. They process every ticket in isolation, forgetting what they did five minutes ago. Keva Brain changes that.
A multi-layered memory system that gives the agent persistent knowledge about your platforms, your customers, and its own past actions.
Traditional AI support agents treat every ticket like a blank slate. A customer writes in about the same billing issue for the third time, and the agent acts surprised. It re-discovers your Shopify store's configuration. It tries the same action that failed yesterday.
Your action_executions table stores every action ever taken — but the agent never sees it. Your customer memory is a 500-character text blob. Your connector has executed 200 actions, but the agent doesn't know what a single one revealed about your platform.
Keva Brain takes all of this data and turns it into intelligence.
Brain data is loaded before the agent runs
No extra latency for your customers. Written after actions complete.
Ticket arrives
A customer submits a support request through email, widget, or API.
Brain loads context
Keva fetches connector knowledge, action history, and runs safety checks. ~500 extra tokens.
Agent acts with memory
The agent uses all 4 layers to make informed decisions. Past actions, platform quirks, safety warnings.
Brain learns from results
After execution, new entries are written to the journal. Every 5th action, connector knowledge refreshes.
Coral steps are the Brain. Everything else is the existing pipeline. Brain operations are non-blocking — if they fail, ticket processing continues normally.
Four layers of memory
Platform config, customer actions, behavioral patterns, and risk signals — each serving a distinct purpose.
Layer 1: Connector Knowledge
“Your Shopify store is running Dawn theme with 340 products.”
Every time Keva takes an action on one of your connected platforms, it builds up knowledge about that specific instance. After every 5th action, an AI summary is generated — capturing your platform's configuration, common patterns, and known issues.
WordPress site running PHP 8.1 with WooCommerce 8.4. Theme: Astra Pro. Product catalog has ~2,400 items. Known issue: bulk product updates occasionally timeout after 30s. Last refreshed Mar 14.
When a new ticket arrives about your WordPress store, Keva already knows your theme, your plugin versions, and that bulk updates need special handling. No discovery phase needed.
Layer 2: Action Journal
“Last week I refunded $49.99 for this customer's order #4821.”
Keva keeps a structured diary of every significant action it takes for each customer. Not just "we helped them" — specific, one-line summaries like "Updated product price from $29 to $39" or "Created DNS CNAME record for api.example.com". Read-only actions (listing, viewing) are excluded to keep it focused.
Past Actions for This Customer: - [Mar 14] Updated product price from $29 to $39 - [Mar 12] Processed refund $49.99 for order #4821 - [Mar 10] Created DNS CNAME record for api.example.com [FAILED] - [Mar 8] Restarted Railway service after deploy failure
When a customer writes back saying "that price change you made last week broke something", Keva knows exactly what it did and when. No guessing, no asking the customer to repeat themselves.
Layer 3: Customer Intelligence
“This customer contacts us about billing issues. Satisfaction is declining.”
Beyond text summaries, Keva tracks structured metadata about each customer. What issues keep coming up? Is their satisfaction improving or declining? Are there risk flags like repeated failures or negative sentiment? This is computed automatically from ticket classifications and action outcomes — no extra AI calls needed.
Recurring issues: billing, shipping_delay, product_inquiry Satisfaction trend: declining Risk flags: negative_sentiment, repeat_issue Total actions: 12 | Last action: Mar 14
Keva can detect that a customer has contacted you about billing three times this month and their sentiment is getting worse. That context changes how it responds — more empathy, faster escalation, proactive offers.
Layer 4: Safety Guardrails
“WARNING: This action type failed 3 times on this connector recently.”
Before the agent proposes any action, Keva checks the action journal for patterns that suggest risk. Has this action type failed recently? Has this specific resource been modified too many times today? Have multiple actions failed for this customer? These warnings appear directly in the agent's context.
Safety Warnings: - WARNING: "shopify.create_refund" failed 2 times on this connector in the last 7 days. - CAUTION: Product #42 was modified 3 times in the last 24 hours. - CAUTION: 4 actions have failed for this customer in the last 7 days. Proceed carefully.
The agent sees these warnings before it acts. If refunds have been failing on your Shopify store, it won't blindly try again — it'll investigate first, or escalate to your team.
Memory for visual actions too
When APIs hit their limits, Keva launches headless automation. The Brain learns from these sessions — extracting reusable playbooks, caching authentication flows, and remembering how to navigate your specific admin consoles.
Learned Playbooks
Builds reusable workflowsWhen Keva successfully executes a multi-step admin workflow, it extracts a reusable playbook. Next time a similar request comes in, it follows the optimized path instead of re-discovering each step.
Cloudflare WAF Rule Update: 1. Authenticate via SSO (cached session) 2. Navigate to Security → WAF 3. Locate managed ruleset by ID 4. Toggle rule sensitivity/action 5. Deploy changes to edge 6. Verify propagation status 7. Confirm via API health check
Session Intelligence
Remembers your platformsAfter a browser session, Keva records what it learned — login flows, menu structures, quirks specific to your setup. This intelligence persists in the action journal alongside API actions.
- Login method (password, SSO, 2FA)
- Navigation path to target
- Selectors that worked
- Actions taken and outcome
- Errors encountered and recovery
The result: Browser automation gets faster and more reliable over time. The first time Keva configures a Cloudflare WAF rule might take 45 seconds. The tenth time, it follows the cached playbook and finishes in 8.
Same agent, completely different behavior
The same AI agent, the same tickets — but with memory, it behaves like a completely different system.
Agent starts from scratch on every ticket
Agent knows your platform config, customer history, and past actions
"Can you remind me what you changed last time?"
Agent references the exact action it took 3 days ago
Blindly retries an action that failed yesterday
Sees a safety warning and tries an alternative approach
Treats a frustrated repeat customer like a first-timer
Detects declining satisfaction and adjusts tone automatically
Discovers your platform setup every single time
Already knows your theme, plugins, and known quirks
Fast, cheap, never in the way
For the technically curious. Keva Brain is designed to be invisible.
Non-blocking
All Brain operations are wrapped in try/catch. If the database query fails or the summarizer errors out, ticket processing continues normally. Zero risk of Brain issues affecting your customers.
Lightweight context
Brain adds a small amount of context to each agent request — platform summaries, recent actions, and safety warnings. Small enough to never impact response speed or cost in any meaningful way.
Template-first summaries
Action journal summaries use pattern matching for common actions (update, create, delete, refund). Only unusual actions trigger an AI call. Most summaries cost zero extra tokens.
Refresh every 5 actions
Connector knowledge is not regenerated on every action — only every 5th one. This keeps AI costs negligible while still maintaining fresh platform understanding.
Auto-pruning
A daily cron job (3 AM UTC) deletes journal entries older than 90 days and resets connector knowledge that hasn't been refreshed in 30 days. Your database stays clean automatically.
Tenant-scoped
Every piece of brain data is scoped to your tenant with foreign key constraints. No data leaks between customers. All queries include tenant_id in their WHERE clause.
Frequently asked
Does Keva Brain slow down ticket processing?
No. Brain data is fetched in parallel with other setup work (browser capabilities, customer memory). The 3 database queries add negligible latency. And if they fail, the agent proceeds without Brain data.
How much does Brain add to my AI costs?
Brain is included in all plans — there's no separate charge. Under the hood, it uses a small amount of AI: connector knowledge is refreshed every 5th action using Haiku (the cheapest, fastest model), and unusual action summaries may trigger a Haiku call. Most action summaries use free template matching. The cost per ticket is fractions of a cent.
Can I see what the Brain knows?
The connector_knowledge and action_journal tables are part of your database. You can query them directly. The Keva dashboard will surface this data in future updates.
What happens if I disconnect a platform?
Connector knowledge is deleted automatically (CASCADE on connector deletion). Action journal entries are retained for historical reference since they're linked to tickets and customers, not the connector itself.
Does Brain share data between my tenants/workspaces?
No. Every piece of brain data is tenant-scoped with a tenant_id foreign key and all queries filter by tenant. There is no cross-tenant data access.
What if the AI generates wrong summaries?
Connector knowledge is regenerated every 5 actions, so incorrect summaries self-correct quickly. Safety warnings are computed from hard data (action outcomes, timestamps), not AI predictions. The worst case for an incorrect summary is a slightly less informed agent — not a wrong action.
An AI agent that actually remembers.
Every ticket makes Keva smarter. Every action deepens its understanding.
Custom setup tailored to your stack. We adapt to you.
Book a Demo