A WhatsApp assistant built around a difficult question: how do you stay useful across many conversations, projects, and changing priorities?
Private beta
“What did I decide about Atlas last time—and what should I do next?”
Illustrative scenario
Remembering a fact is only the start. The useful part is finding it later, connecting it to the right request, and letting newer information update the picture without mixing unrelated work.
RecallRelevant memories
→RouteFast reply or research
→RespondOne consistent voice
ExtractFacts and actions
→ReconcileDuplicates and updates
→Retrieve laterContext when relevant
A branching workflow, with memory maintained alongside the conversation.
Semantic search retrieves relevant memories within the user’s namespace, with rules loaded separately. Recent context complements retrieved matches, so useful older information has a path back into the conversation.
Decision: retrieve relevant context instead of replaying the entire history.
The orchestrator can move from recall to a reply, use web search, or take a deeper research path. Short requests can stay short. Tool use follows the work required rather than one mandatory sequence.
Decision: spend latency and model calls selectively.
Background or post-response extraction identifies memories, rules, and structured records such as actions. Central writes attach source, confidence, scope, and entity metadata. Similarity checks skip near-duplicates and identify likely updates.
Decision: maintain memory as data with provenance.
Reasoning and tool results feed the response layer, which handles Edith’s voice before delivery through WhatsApp. Transport, reasoning, and presentation have separate responsibilities.
Decision: keep a consistent experience across different workflows.
Select a stage to explore the engineering.
A user correction should carry more authority than an inferred fact. I centralized memory writes so inferred content cannot overwrite user-sourced information, added semantic deduplication, and invalidated cached context after changes.
Similarity is a useful signal, but it is not proof that two facts contradict each other. I also benchmark routing with the actual runtime prompt: the selected baseline passed 18 of 21 repeated cases. A faster candidate reached 15 of 21, so I kept the stronger routing baseline and targeted other layers for cost savings.
2.18s median orchestration latency · April 4, 2026
Repeated benchmark cases using the runtime prompt; measures routing, not complete reply latency.
WhatsApp Cloud API · Agents · Vector retrieval · DynamoDB