WORKSHOP · DEEP DIVE · WS·02

Enterprise AI Memory Architecture

Three layers. Three jobs. One rule: every fact has exactly one home. 13 slides, 40 minutes.

13 SLIDES · ~40 MIN · PLAYBOOK OVERVIEW
Key Takeaways
  • One fact, one home. The moment you mirror facts, you've built a museum of contradictions.
  • $1M — annual cost of memory failure at a 100-person firm. Math: 30 min/day × $80/hr blended × 100 people × 250 working days.
  • ~30 min — per knowledge worker per day spent re-establishing context (assumes $80/hr mid-market blended rate).
  • 4M tokens — per day re-establishing context at ~20 AI interactions per person — roughly 6,000 pages of text, or ~$20/day in inference at current rates.
  • Map your firm's knowledge into the three types — relational, operational, documentary — before you touch a tool.
WORKSHOP · DEEP DIVE

Enterprise AI Memory Architecture

Three layers. Three jobs. One rule — every fact has exactly one home.

BuildClub Academy
01 · WHY THIS MATTERS

A 100-person firm loses ~$1M/year to AI that doesn't remember.

$1M
annual cost of memory failure at a 100-person firm. Math: 30 min/day × $80/hr blended × 100 people × 250 working days.
~30 min
per knowledge worker per day spent re-establishing context (assumes $80/hr mid-market blended rate).
4M tokens
per day re-establishing context at ~20 AI interactions per person — roughly 6,000 pages of text, or ~$20/day in inference at current rates.

A bigger context window is not a memory strategy. It is a way to pay more for the same forgetting.

02 · THE THESIS

One fact, one home.
The moment you mirror facts,
you've built a museum of contradictions.

The Thesis
03 · THE KNOWLEDGE TYPES

Enterprise knowledge is three things — and one layer can't serve all three.

INPUT SOURCES KNOWLEDGE TYPES Meeting transcript Chat message Decision memo Process config Final PDF Client-received doc TYPE 01 Relational What people said, decided, believe. Conversations · calls · meetings TYPE 02 Operational How the firm works. Standing decisions. Configs · preferences · process state TYPE 03 Documentary Verbatim source material. Final PDFs · contracts · deliverables

Each input naturally belongs to one type. The architecture honors that. A memory system that flattens them will fail at all three.

04 · THE ARCHITECTURE

Three layers, one rule, one source of truth per knowledge type.

QUERY ORCHESTRATION LAYER The agent, app, or workflow that fans out to the right layer and synthesizes the answer. 1 LAYER 1 Temporal Knowledge Graph Zep RELATIONAL KNOWLEDGE Conversations, decisions, entities, time-stamped facts. INPUT Meeting transcripts 2 LAYER 2 Portable Memory Store Mem0 OPERATIONAL KNOWLEDGE Atomic facts about how the firm works. Scoped IDs. INPUT AI session flushes 3 LAYER 3 Vector Document Store pgvector DOCUMENTARY KNOWLEDGE Verbatim text. Citation-grade retrieval from final documents. INPUT Curated documents

Three independent stores. One orchestration layer. Three rules, three jobs — and exactly one home for every fact.

05 · LAYER 1 · ZEP

Temporal knowledge graph. The conversation memory of the enterprise.

1
Bi-temporal storage
Every fact has event-time AND ingestion-time. The graph remembers what we used to think.
2
Auto-extraction
Hand it raw conversation text. Entities, facts, relationships, and invalidations emerge automatically. No ontology to maintain.
3
One API call
thread.add_messages() appends to the thread AND ingests into the graph. No sync bugs because there is nothing to sync.
BI-TEMPORAL — A FACT REMEMBERS WHEN IT WAS TRUE AND WHEN WE LEARNED IT
FACT: "ACME RENEWS AT $1.2M" EVENT TIME When the fact was true in the world opens superseded ...new fact: "Acme renews at $1.8M" INGESTION TIME When the system learned about it learned v1 learned v2 TODAY May 1 May 5 May 25 Jun 1 Today

If the question starts with "what did we say about…", this is the layer that answers it.

06 · LAYER 2 · MEM0

Portable memory store. Atomic facts. Scoped IDs. Vendor-independent.

1
Atomic facts
One memory = one durable, self-contained fact. No paragraphs.
2
Scoped IDs
firm_ops, client_{name}, project_{code}. Reuse a scope and you contaminate the store.
3
Broad-to-specific tags
First tag = canonical entity. Last tag = the precise fact category. Three rules and the store stays usable for years.
4
Two feeds
Nightly flush from AI sessions. Optional graph export when portability outweighs duplication cost.
SCOPED IDS + TAG CHAINS — HOW MEM0 STAYS ORGANIZED AT SCALE
MEM0 MEMORY STORE one fact, one home SCOPE firm_ops SCOPE client_acme SCOPE project_horizon firm_ops prefs format three_columns → atomic fact stored at this address firm_ops process invoicing net_30_2pct → atomic fact stored at this address client_acme contact primary mary_chen_cfo → atomic fact stored at this address client_acme prefs meeting tue_2pm_pt → atomic fact stored at this address proj_horizon status phase launching_q3 → atomic fact stored at this address proj_horizon decision scope no_mobile_v1 → atomic fact stored at this address

Vendor memory is a hostage situation. Mem0 is how you keep your firm's accumulated decisions when you switch AI tools next year.

07 · LAYER 3 · PGVECTOR

Vector document store. Verbatim retrieval. Citation-grade.

1
Curate or skip
A vector store full of drafts is a search engine that retrieves the wrong answer with high confidence.
2
Chunking is destiny
A great embedding on bad chunks loses to a mediocre embedding on great chunks. Every time.
3
Rerank or miss
Vector search picks the neighborhood. Reranking picks the right house. Skip it and you retrieve close-but-wrong.
4
n8n runs it, AI built it
Use the AI to build the pipeline. Use n8n to run it. Tokens are for thinking, not for plumbing.
INGESTION PIPELINE — PARSE · CHUNK · EMBED · STORE · RETRIEVE · RERANK
DOCUMENT INGESTION PIPELINE Each step transforms what the previous step emitted. Arrows show the data shape. 1 Parse LlamaParse → markdown 2 Chunk Python → tokens 3 Embed Voyage AI → vectors 4 Store pgvector → index 5 Retrieve cosine → candidates 6 Rerank rerank-2.5 → ranked chunks n8n runs every node. The AI built it; n8n executes it in production.

If the question requires a citation, this is the layer that answers it.

08 · THE FILING SYSTEM

The one-sentence rule that keeps Layer 3 clean.

1
The protocol
When it's final, drop it in the KB folder. That is the entire human-facing protocol.
2
The shadow
Drafts live in working folders. The KB folder only sees _FINAL and client-received documents.
3
The audit
Auto-ingest client-received PDFs — they will not be drafts, they will not be junk, and they will save you in every dispute.

Layer 3 is only as good as the discipline you teach the team.

09 · THE QUERY FLOW

Three layers, queried in parallel. The user asks one question.

USER QUESTION "What did we decide about Acme's renewal?" 1 Zep Relational ~600 MS 2 Mem0 Operational ~200 MS 3 pgvector Documentary ~400 MS SYNTHESIZED ANSWER Total wall-clock: ~700 ms (parallel)

Three layers, queried in parallel, synthesized in one. The user asks one question; the architecture quietly does the work of three.

10 · COSTS AND ROI

~$700/mo at small scale. ~$7K–16K/mo at enterprise. Pays for itself in weeks.

MONTHLY OPERATING COST ~$700 SMALL FIRM 10–20 knowledge workers $7K–16K ENTERPRISE 100+ knowledge workers ≈ cost of one mid-level analyst SCALES LINEARLY ~10× users · ~10× cost
Takeaways
  • 1 At enterprise tier, this is approximately the loaded cost of one mid-level analyst.
  • 2 The architecture pays for itself in single-digit weeks if the firm has more than 100 knowledge workers.
  • 3 Linear scaling: cost per user falls as headcount rises, while the value of cross-team continuity rises faster.

Cost scales roughly linearly with users. The math gets better, not worse, as the firm grows.

11 · IMPLEMENTATION SEQUENCE

Twenty weeks. One layer at a time. Discipline before scale.

IMPLEMENTATION SEQUENCE · 20 WEEKS Highest-ROI layer first. Discipline before scale. Layer 3 last because it's the most demanding to curate. Phase 1 — Layer 1 (Zep) WK 1–4 First transcripts in graph Phase 2 — Layer 2 (Mem0) WK 5–8 Nightly flush working Phase 3 — Filing System & Taxonomy WK 9–12 KB folder discipline live Phase 4 — Layer 3 (pgvector) WK 13–20 Citation-grade retrieval wk 0 wk 4 wk 8 wk 12 wk 16 wk 20 One layer at a time. Each phase ships a usable thing before the next begins.

Highest ROI first (Layer 1). Discipline next (filing system). Heaviest curation last (Layer 3). The order is not optional.

CLOSING

Five things to do Monday.

Monday Morning
  1. 1
    Map your firm's knowledge into the three types — relational, operational, documentary — before you touch a tool.
  2. 2
    Start with Layer 1 (Zep). Turn on Fireflies and ingest one engagement's transcripts. Highest ROI per unit effort.
  3. 3
    Establish the one-sentence rule: when it's final, drop it in the KB folder. Train the team before you build Layer 3.
  4. 4
    Build the prototype in the smart, expensive place. Run it in the dumb, cheap place. Use n8n for plumbing.
  5. 5
    Refuse mirroring. One fact, one home. Make it the team's discipline before the architecture compounds it.
WS·02 · Questions CEOs Ask

Frequently Asked Questions

What is the core idea of Enterprise AI Memory Architecture?
One fact, one home. The moment you mirror facts, you've built a museum of contradictions.
What does the data say a CEO should pay attention to?
$1M annual cost of memory failure at a 100-person firm. Math: 30 min/day × $80/hr blended × 100 people × 250 working days. ~30 min per knowledge worker per day spent re-establishing context (assumes $80/hr mid-market blended rate).
What should a CEO do Monday morning after reading Enterprise AI Memory Architecture?
Start here: Map your firm's knowledge into the three types — relational, operational, documentary — before you touch a tool; Start with Layer 1 (Zep). Turn on Fireflies and ingest one engagement's transcripts. Highest ROI per unit effort; Establish the one-sentence rule: when it's final, drop it in the KB folder. Train the team before you build Layer 3.
Where do the claims in Enterprise AI Memory Architecture come from?
The playbook cites BuildClub Ops Research — Enterprise AI Memory Architecture Reference (v1.0, May 2026); BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §1.4 (May 2026); BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §2.3 (May 2026); BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §1.2 (May 2026).
WS·02 · Sources

All Sources in This Playbook

  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference (v1.0, May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §1.4 (May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §2.3 (May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §1.2 (May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §2.1 (May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §3.2–3.4 (May 2026)
  • Zep — Temporal Knowledge Graph documentation
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §4 (May 2026)
  • Mem0 — Open-source memory layer for AI agents
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §5 (May 2026)
  • Supabase pgvector — Postgres vector extension
  • Voyage AI — embeddings and rerank-2.5
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §6 (May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §7.4 (May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §8.1–8.4 (May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §9.1 (May 2026)
  • BuildClub Ops Research — Enterprise AI Memory Architecture Reference, §9 + §10 (May 2026)
00 / 13