Architecture · Future State

Connected Vault Architecture

Data-set-oriented vault design with agent memory layer, access-controlled data boundaries, and time-aware knowledge management. Evolution of the current vault system toward a fully autonomous, self-maintaining intelligence platform.

Future State Architecture Access-Controlled Data Boundaries Microsoft Foundry Memory Layer
Status as of 2026-05-02: Architecture built and producing value across the org. Memory Layer (Microsoft Foundry) is the only outstanding capability — implementation planned for the coming weeks. IT has formally blessed the system to scale company-wide and to take on data integrations into the architecture.

Adoption: 4 active users (up from 3 baseline in Apr 2026); 10+ colleagues self-pulled in with specific use cases. CEO + exec sponsorship in place. CRO called the initiative "brilliant" (Teams DM, May 2026).
The Layer Stack · top → bottom

01Data Set Vaults — Agent-Only Reads

Raw data stays behind explicit access boundaries. Most humans never read these directly. Agents are the only readers; they synthesize PII-free, sanitized context into the world model.

HubSpot

Contacts, leads, deals, sales pipeline, account data.

Agent-only · live API

Jira

Sprint stats, dev tickets.

Agent-only · live API

Product Docs

Official product docs, shipped feature documentation.

All-org read · external-owned

Product

Product landscape, relationships, 200+ product records.

All-org read · non-sensitive

Customer Deployments

Customer servers, configs, setup state, fleet list.

Agent-only · ops-team admin

User Research

Interviews, usability studies, surveys, insights library.

Design + agents · PII-cleared

Transcripts

Meeting recordings, anytime, any format.

Agent-only · parser-only

Analytics

Usage telemetry, feature adoption rates.

Agent-only · planned

Support

Support tickets, issue patterns.

Agent-only · CS admins

Corporate Policies

HR, IT, legal, compliance docs.

Agent-only · reference-only

02Knowledge Standards knowledge-standards

Shared primitives — metadata schema, tagging taxonomy, writing standards, agent configs, memory layer schema, playbooks. Versioned, distributed via submodule. Tier 1 branch protection — all changes require PR + review. Acts as the contract that prevents garbage-in, garbage-out across the system.

03CEO Vault — Strategic Intent (Dual-Zone)

Top-down strategic direction. Highest-authority input to the system — sets the frame all other layers operate within.

Published Zone

Feeds world model · readable by agents

  • Strategic Priorities — top-down direction, updated quarterly
  • CEO Decisions — org-level choices with rationale; supersedes conflicting team decisions
  • Constraints & Guardrails — what we will NOT do; budgets, timelines, non-negotiables
  • Open Questions — CEO wants answered; agents prioritize these signals

Private Zone

Isolated · never enters shared state

  • Board prep
  • Sensitive strategy
  • Personnel
  • M&A
  • Compensation
  • Draft thinking before publishing

CEO published intent becomes highest-priority context in the world model — agents weight it above bottom-up signals.

04Shared World Model world-model

This is what people actually access. Synthesized, PII-free context produced by agents from restricted data sets. No raw customer data, no ticket PII, no individual names from transcripts — only patterns, counts, themes, and decisions. Fed by agents (bottom-up) and CEO intent (top-down). Tier 1 branch protection.

Live Operational State agent-maintained

  • current-state.md — active priorities, open questions, what changed this sprint
  • active-dris.md — who owns what problem, time-boxed with expiry
  • decisions-log.md — cross-org decisions mirrored from all vaults
  • blockers.md — what is stuck, severity-rated
  • product-health/ — one file per product: status, blockers, milestones

Synthesized Insights agent-generated

  • Distilled Insights — cross-dataset patterns, agent-synthesized findings
  • Customer Picture — HubSpot + Customer Deployments + Support synthesized per customer

05Agent Memory Layer — Microsoft Foundry planned · in implementation

Structured working memory for agents. A scratchpad of current state, evolving facts, lessons, and decision history. Records have types, metadata, and decay rules. Not a vault — a cognitive substrate.

Record Types

state

Current truth — what's the situation right now.

Decays in 30d without reinforcement.

lesson

Historical truth — patterns/insights that persist indefinitely.

No decay.

decision

Specific choice with rationale; supersedable, forms a chain.

No decay; supersedable.

event

Something that happened; immutable historical record.

Immutable.

Metadata Schema

created_at · valid_from / valid_until · confidence · importance · reinforcement_count · last_reinforced_at · superseded_by · tags

Retrieval Scoring (weights per kind)

state

Recency heavy · tag-match heavy · confidence heavy. Superseded records killed, never surfaced.

lesson

Importance heavy · tag-match heavy · recency ~0. Old lessons just as valid as new.

decision

Importance + tag match. Returns current + full supersession chain so the agent sees evolution.

event

Tag-match heavy · importance · recency ignored. Founding event from 5 years ago = yesterday's event.

Decay & Irrelevance

06Intelligence Layer — Agents

Read from data set vaults (per access grant). Write to world model and memory layer. Each agent has an explicit data-set access list.

Active

meeting-parser decision-logger daily-briefing user-story-writer scope-guard ux-flow-reviewer

New

world-model-maintainer cross-vault-synthesizer insight-tracer

07Team Vaults

Where work-in-progress lives. Agents extract knowledge upward; teams write here; agents synthesize to world model. Tier 2 branch protection — force pushes and deletions blocked, direct commits to main allowed.

software-product-design

Decisions, work tracking, research, design files.

Active

software-product-management

Roadmaps, requirements, metrics, decisions.

Active

Engineering

ADRs, runbooks, system specs, on-call.

Planned

Customer Success

Customer health, escalations, playbooks.

Planned

08Personal Vaults — Edge Nodes

Sensitive and scratch content. Never flows into shared state. Cryptographically isolated by design.

Director Vault

Team, sensitive, hiring, org strategy, career.

Individual Vaults

Working notes, drafts, career, daily notes.

09People at the Edge

Design · PM · Engineering · Customer Success · Leadership · New hires.

Intelligence delivered — no more asking someone for context. Read the world model directly. Get current state without asking anyone. Agents read the memory layer for evolving context. daily-briefing delivers a compiled view each morning.

Security Model

Twelve enforcement mechanisms

What makes this an architected intelligence platform, not a knowledge base.

01Agents as Sanitization Layer

Most humans never read raw data set vaults; agents are the controlled intermediary. They read raw data (PII, operational details, verbatim transcripts) and synthesize sanitized, anonymized outputs into the world model. A support ticket with a customer name becomes "ticket category frequency count" — the person reading the world model never sees raw data they shouldn't have access to.

02Per-Dataset Access Control

Each data set vault has an explicit consumer list defining which agents and which human roles (if any) can read it.

03Personal Vault Isolation

Personal vaults are cryptographically edge nodes. Content never enters shared state, world model, or memory layer. No agent reads personal vaults.

04Write Direction Enforcement

Data set vaults are read-only to agents. Agents can write to: world model (synthesis), memory layer (working state), team vaults (outputs). They cannot modify source data. World model is append-only for agents.

05Memory Layer Scoping

Memory records inherit the access scope of the data set they were derived from. A memory tagged "hubspot" is only surfaced to agents with HubSpot access. Cross-dataset memories require explicit dual-access.

06Archive, Never Delete

Decayed or superseded records are archived, not deleted. Full audit trail preserved. Archived records remain queryable via explicit archive search but are excluded from normal retrieval.

07Supersession Chain Integrity

When a decision supersedes another, the full chain is preserved. Agents see the evolution: current decision plus what it replaced and why. No decision is silently overwritten. Rationale persists.

08Confidence Decay as Safety

State records automatically lose confidence. If no agent or human reinforces a fact within 30 days, it degrades toward archive. This prevents the system from asserting outdated information as current truth.

09Knowledge Standards as Contract

knowledge-standards defines the schema all data must conform to. Any record without valid metadata is rejected. Tags must come from the approved taxonomy. Prevents garbage-in, garbage-out.

10CEO Vault Dual-Zone Isolation

The CEO vault has two cryptographically separate zones. The published intent zone is readable by the world-model-maintainer agent only — it writes CEO priorities into current-state.md. The private zone is fully isolated: board prep, M&A, personnel, and compensation never enter any shared layer. Publishing is an explicit, deliberate act by the CEO.

11Top-Down Authority Hierarchy

CEO decisions in the published zone carry a system-wide authority flag. When a CEO decision conflicts with a team-level decision, agents surface both but mark the CEO decision as authoritative. Team decisions are not silently overwritten — they are flagged as potentially superseded, requiring the team to acknowledge the CEO direction.

12Branch Protection Tiers

Two protection tiers enforced via GitHub Rulesets. Tier 1 (world model + knowledge standards): restrict deletions, restrict force pushes, require pull request with 1 approval before any merge to main. Tier 2 (team vaults): restrict deletions and block force pushes — direct push allowed to keep daily workflows fast. Personal vaults: no branch protection, owner-only repos.