Skip to main content

[ANCHOR] Qontak | AI Agent | AI Agent Memory

This is an ANCHOR PRD. It is the master index for the AI Agent Memory initiative — it orchestrates the phases beneath it and never describes a single delivery. Per-phase execution detail (constraints, stories, dependencies, observability) lives in the child Phase PRDs linked in the Phase Index.

Theme 1: Make the Intelligence Promise Real | Objective 1A: Strengthen Agentic AI Capabilities | Q4 2026 | Owner: Dimas Fauzi Hidayat

HEADER BLOCK

FieldValue
PMDimas Fauzi Hidayat (Product Manager, Mekari Qontak)
PRD Version1.0
StatusDRAFT
PRD TypeANCHOR
EpicTBD — created at grooming
SquadBOT — Chatbot Squad
RFC LinkN/A — anchor carries no delivery; Phase-1 RFC pending (gated on the DSAI contract + CEBE RFC)
Figma MasterN/A — Phase 1 has no UI surface (flag-only rollout)
AnchorYes — this IS the Anchor
Labelsepic:qontak-chat | module:ai-agent | feature:ai-agent-memory
Last Updated2026-07-22

Table of Contents


Phase Index

PhaseGoalPRD LinkEpic KeyStatusShipped
Phase 1 — Q4 2026: Read-side CEBE memoryThe autonomous AI Agent reads the customer's cross-module CEBE profile and history (keyed qontak_customer_id) at conversation start and uses it in responses — org-flag-gated, with graceful degradation whenever CEBE data, identity resolution, or the latency budget is unavailable.phase-1-read-side-cebe-memory.mdTBDDraft

Future considerations — deliberately NOT committed phases (revisited only after Phase 1 ships and the CEBE event schema stabilizes):

  • Agent-authored memory write-back — the agent emits learned facts (preferences, commitments, declared attributes) back into CEBE. Requires net-new AI event types in the standardized event schema, so it must be negotiated into the CEBE schema roadmap with the CDP squad first.
  • Tenant memory controls — admin UI to view/edit/delete a customer's memory, retention settings, and opt-out; carries the UU PDP erasure obligations that read-side consumption defers to CEBE's own governance.

Backbone dependency: every phase of this initiative consumes Customer Event-Based Engine (CEBE) (TF-3302, CDP squad, DRI Zhelia Alifa). The CEBE ANCHOR names this initiative as its flagship downstream consumer: roadmap row "Agentic AI reads CEBE memory — Eng (Chatbot) — Q4 2026".


2. One-liner + Problem

One-liner: Give the autonomous AI Agent customer-level memory by reading the customer's cross-module CEBE history before responding — recognizing returning customers instead of answering context-blind.

Problem: The AI Agent is context-blind across modules and stateless per turn. On the chatbot BE only the current message is sent to the AI service (send_message_with_resolve.rbquestion: last_history&.input_content); conversation state lives in the AI-service thread keyed by room_id/conversation_id and evaporates with the conversation. No cross-conversation or contact-level memory exists anywhere in the stack. The CEBE ANCHOR states the business consequence precisely: "a ticket resolved in the support module never informs the AI Agent handling a follow-up sale." A returning customer is greeted as a stranger, re-asked facts the business already holds (open tickets, past orders, loyalty tier, campaign opt-outs), and an at-risk or high-value customer is indistinguishable from a new one — capping containment and CSAT, the exact gains the Intelligence theme promises. The cross-vertical CX-journey research (6-vertical action-plan workshop, 2026-03-11) independently converged on the same mechanic: CDP-driven personalization with CDP as identity resolution recurs in every vertical's target journey — this initiative is that mechanic made real for the AI Agent.


3. What Happens If We Don't Build This

  • CEBE loses its flagship consumer. The CEBE Q4 roadmap row "Agentic AI reads CEBE memory" and its success metric (AI Agent uses CEBE memory in ≥50% of conversations by end of Q4 2026) have no owner and silently slip — weakening the case for the backbone the whole Intelligence theme builds on.
  • The agent stays context-blind. Containment and CSAT stay capped at what a memoryless agent can do; returning customers keep being re-asked what the business already knows.
  • Cross-module data access gets re-solved ad hoc. Without the CEBE consumer path, future agent-context needs get wired as point-to-point lookups from the chatbot BE — the exact per-squad re-payment CEBE exists to eliminate.
  • The personalization gap widens. Every target journey in the 6-vertical CX research assumes CDP-driven personalization; competitors shipping agent memory make a memoryless agent a visible competitive deficit in deals.

4. Strategic Context (Optional — not scored)

Module Positioning

Chatbot & AI → AI Agent (Generative AI). Memory is the layer that converts CEBE's unified customer data into agent behavior — the first of the CEBE downstream consumers (segmentation, marketing automation, health score are siblings owned by other squads). Within the chatbot portfolio it is orthogonal to AI Agent Knowledge: Knowledge = company-level (what the business knows), Memory = customer-level (what we know about this customer).

Pricing & Package Implications

Rides the existing autonomous-agent entitlement (Plus / Ultimate / 360) — no new SKU proposed for Phase 1. Whether memory later becomes a packaging differentiator (e.g. higher-tier-only) is an open question (§8) to resolve with pricing before GA.


5. Target Users + Persona Context

PersonaRoleGoalPainWorkaround
Primary — Business Admin / CS LeadOwns the org's production AI Agent and its containment/CSAT outcomesAn agent that recognizes returning customers and answers with full account context, lifting containment without more human agentsThe agent re-asks known facts and treats every conversation as first contact; customers escalate out of frustrationWrite ever-longer static instructions/KB entries; route "known customers" to humans; accept the CSAT cap
Secondary — AI Agent (system actor)The autonomous agent acting on the business's behalf in live conversationsRead the customer's cross-module profile/history before respondingReceives only the current message + thread; no access to tickets, orders, campaigns, loyaltyNone — the data exists in silos the agent cannot reach
Tertiary — End CustomerThe business's customer chatting on WhatsApp/webchatContinuous service — not repeating their history every conversationRepeats context each session; gets generic answers despite a long relationshipAsks for a human agent

6. Success Metrics

CategoryMetricDefinitionBaselineTarget
Adoption & Usage⭐ AI Agent uses CEBE memory% of AI-handled conversations where the agent reads CEBE history before responding (inherited from the CEBE ANCHOR — this initiative delivers it)0% (no memory exists)≥ 50% of conversations by end of Q4 2026
Quality & ImpactContainment delta (memory vs. non-memory)RESOLVE_AI share of AI-handled rooms, memory cohort vs. control (spine: rooms.closed_reason)Cohort baseline read at rollout+3 pp within 60 days of GA (target reviewed at baseline read)
Quality & ImpactCSAT deltaAvg CSAT on memory-cohort conversations vs. control, where CSAT is collectedCohort baseline read at rollout+0.2 within 60 days of GA
Efficiency & GuardrailAdded response latencyp95 latency added to the agent's first response by the CEBE read (subsequent turns: 0 — context is cached in the thread)N/A — newp95 ≤ 2s; breach alerts and auto-skip (guardrail, not a growth metric)

Instrumentation and per-response telemetry are Phase-1 scope (see the phase PRD's Observability section) — the ⭐ metric is unmeasurable without them.


7. Key Decisions + Alternatives Rejected

7a — Decisions Made

DateDecisionRationale
2026-07-22Stand up a new initiative with its own ANCHOR (not a phase under autonomous-ai-agent)Memory spans multiple quarters and two external squads (CDP, DSAI) with its own interfaces and metrics; it needs its own epic, roadmap row, and dependency edges
2026-07-22v1 is read-side only — the agent reads CEBE; it writes nothing backFastest path to the CEBE Q4 consumer commitment and the ≥50% metric; write-back needs net-new AI event types in the standardized schema, which is a CDP-squad schema negotiation, not a chatbot deliverable
2026-07-22DSAI queries the CEBE retriever directly at runtime (chatbot-BE-side assembly rejected)Context assembly belongs beside the reasoning engine: DSAI owns the thread state and prompt assembly, so injecting via the chatbot BE would push customer context through the thread-message contract, duplicate retriever-consumption logic in a second squad's runtime, and add a hop on the reply path. Avoidability check run per the cross-team dependency rule — the squad-owned alternative was evaluated (7b row 1) and rejected with eyes open; the DSAI dependency is accepted with a hard contract/commitment gate before BUILD (Phase-1 PRD, Dependencies)
2026-07-22Phase 1 ships flag-only, with no FE surfaceNo memory design exists in the design system prototype (design census 2026-07-22); an org-level config toggle would gate Q4 on a design dependency for zero adoption benefit during staged rollout

7b — Alternatives Rejected

AlternativeWhy RejectedDate
Chatbot BE assembles context (BE queries CEBE, injects via the predict payload / SendContext thread push)Duplicates CEBE consumption logic outside the reasoning engine, splits ownership of what the model sees, and adds a hop on the customer reply path. Known trade-off: DSAI delivery has a slow/unreliable precedent — mitigated by the contract gate, not by re-homing the work2026-07-22
Reuse omnichannel_room_summaries as memory v1Per-conversation and chatbot-GPT-scoped — not cross-module, and does not deliver the CEBE consumer commitment; remains a possible future additional source2026-07-22
Wait for write-back event types and ship read+write togetherCouples Q4 delivery to an unscheduled CEBE schema extension; read-side alone already delivers the ≥50% metric and all Phase-1 value2026-07-22

8. Open Questions

#TypeQuestionOwnerDeadline
1RiskCEBE core infra slips. The joint Eng+Data RFC (due Jul 2026) is escalated in the CDP sprint; a Q3 slip cascades into this initiative's Q4 target. Mitigation: inject this initiative's query/latency/auth requirements into the CEBE RFC now (Phase-1 PRD, Dependencies); phase-gate BUILD on CEBE staging availability; fallback = descope memory content to whichever modules have connectedDimas Fauzi Hidayat + Zhelia Alifa2026-08-15
2RiskDSAI capacity/commitment for Q4 not yet secured (retriever client + context injection are DSAI-side work). Mitigation: written contract + committed timeline is a BUILD gate — no RFC-approved build starts without itDimas Fauzi Hidayat + DSAI lead2026-08-31
3Assumptionqontak_customer_id is resolvable for enough AI-handled conversations to reach the ≥50% adoption target (identity-coverage feasibility — today the predict payload carries only channel_contact_id)Chatbot BE + CDP squad2026-08-15
4Open QuestionPricing/packaging: does memory stay bundled in existing autonomous-agent tiers, or become a tier differentiator later?Dimas Fauzi Hidayat + Pricing2026-09-30

9. User Stories & Acceptance Criteria

ANCHOR PRD — stories delegated to the phase PRDs. This is an initiative-level anchor; its detailed, testable user stories and acceptance criteria live in the per-phase PRDs listed in the Phase Index above. The anchor tracks only the initiative-level outcome below.

[MEM-E01] Initiative epic — As a business running the Qontak AI Agent, I want the agent to remember and use my customers' cross-module history, so that returning customers get continuous, context-aware service instead of starting from zero in every conversation.

IDAcceptance criterion
AC-1Every phase PRD in the Phase Index reaches READY and its own acceptance criteria are met, delivering the initiative-level outcome in Success Metrics (⭐ ≥50% of AI conversations read CEBE memory by end of Q4 2026).

PRD CHANGELOG

VersionDateBySectionTypeSummary
1.02026-07-22Dimas Fauzi Hidayat / ClaudeAllADDEDInitial ANCHOR: initiative framing as CEBE's flagship downstream consumer, Phase Index (Phase 1 read-side), north-star metrics, DSAI-direct decision + avoidability record.