AI · Multi-agent · Under NDA
A memory layer four AI agents share.
A global advertising holding company had institutional knowledge scattered across people, tools and threads. Edenic engineers built a governed organizational memory system — four specialist agents working across chat, Slack and email on one semantic memory layer, with deterministic conflict resolution and a human in the loop on every low-confidence write.
specialist AI agents
surfaces: chat · Slack · email
reusable service modules
sprints to delivery
Snapshot
Client
A global advertising & media holding company (London) — under NDA
The problem
Institutional knowledge scattered across people, tools and conversations
The system
One governed, queryable memory layer every AI agent reads and writes through
Governance
Human review on every low-confidence write, with a full audit trail
The pain
- Knowledge lived in people's heads, threads, and docs — not anywhere queryable
- AI agents answered from stale or contradictory facts with no way to tell which won
- No audit trail for what an agent knew, when it knew it, or who approved it
- Every new surface (chat, Slack, email) meant rebuilding the same plumbing
Objectives
- One authoritative memory layer shared by every agent and surface
- Deterministic conflict resolution — no LLM arbitration deciding what's true
- Human-in-the-loop approval before low-confidence facts become authoritative
- Full telemetry and an audit trail on every decision the system makes
Surfaces
Orchestration
Specialist agents
Memory & data
Integrations
Client architecture is confidential — this is a sanitized, high-level redraw with agent names and internal systems generalised.
What we built
Hybrid memory layer
Vector search handles semantic recall; SQL handles deterministic lookups, config and audit. Each store does what it is actually good at.
Deterministic conflict resolution
When two facts contradict, a three-tier rule decides — recency, then confidence, then source authority. Losers are deprecated so stale data never resurfaces.
Draft queue + human review
Writes missing a source, a date, or enough confidence are demoted to draft and queued for a human to approve, correct, or reject.
Four specialist agents
An orchestrator delegates to research, task, and content agents in parallel, then synthesizes one coherent answer.
Runtime prompt assembly
Agents are governed by retrieved knowledge, not hardcoded instructions — memory, client context and skills are assembled per request within a token budget.
Nightly consolidation
A batch job archives stale and low-confidence memories and emits a health report — cheaper and more auditable than recalculating on every write.
Key takeaways
- One memory layer replaced knowledge scattered across people and tools
- Deterministic resolution beat LLM arbitration — auditable, repeatable, no hallucinated truth
- Human-in-the-loop kept quality high without blocking the whole write path
- Layered architecture meant a new surface plugged in without touching the agents
Building something agentic?
This is the calibre of engineer Edenic sources, vets, and places. Talk to a founder.
Talk to a founder ↗