Beever Atlas — Adrian Leung

Enterprise AI memory layer ·

Designing the memory an AI agent is allowed to have.

ProductGTM

#453
of 18,000+ MCP servers on PulseMCP: 6
AI-agent discovery registries listed

Overview

Role: Product Manager
Timeline: 2026.05 — present
Platform: Web App · MCP · AI Agents
Scope: v2 direction, GTM

Beever Atlas turns scattered team chats, notes, and documents, into a permanent, cited AI memory center and LLM wiki. I joined after the CTO and an AI engineer had already built v1 and shipped it. GTM, the v1→v2 direction, and the UX/brand work.

Everything Atlas reads from — Slack, Teams, Google Drive, Jira, and email landing in v2.

Sources

Slack · Teams · Drive · Jira

Ingest & cite

Every fact keeps its source

Memory spaces

Semantic + graph, RBAC applied

Living wiki

Pages recompile themselves

Agents over MCP — AI agents read and write the same memory

Target user

The teammates who run their own agents

AI-literate

Delegates day-to-day ops to their agents.

Context work

Copies threads, docs and decisions before agents can act.

Cross-domain work

Across marketing, product, sales, etc.

Teammates reviewing an AI agent-generated proposal together at a shared desk

Situation

Every tool had its own way of saving knowledge — and v1's real flaw was a security one, not a hypothetical.

Conversational knowledge loss

Knowledge points and key dicision makings scattering in different locations, different formats(e.g. Audio, images, pdf, and notes).

Ambient authority — v1's real flaw

v1 used one field to do four jobs at once, including access control. Anyone who could reach the bot could ask about any synced channel.

Diagram of where a teammate's context lives before Beever Atlas — Where a teammate's context actually lived before Atlas — scattered across tools, none written for an agent to read.

User side

Everyone use diverse tools and formats. They aren't easy for an agent to pick up.

Votee side

v1 was AI research-initiated and the product shipping journey is unclear.

Six tools, six silos

Every tool remembers differently. None of it connects.

Slack

Starred message

in-app only

Linear

Thread summary

in-app only

Gmail

Labelled email

in-app only

Docs

Meeting notes

not searchable

Calls

Audio recording

not searchable

Screens

Screenshot

not searchable

Chat & email: searchable in-app only. Notes, audio, screenshots: not searchable anywhere.

No benchmark fit the case

We had no quantitative metrics to show clients why our memory layer was actually better — no existing AI industry benchmark measured a cited, permission-aware agent-plus-memory system in our case.

The memory pipeline

Eight agents read the stream. Nobody hand-maintains a page.

Injection agents

8 small agents read the stream and file every fact — atomic, cited, and permission-stamped.

Dreaming stage

Periodically revisits the wiki to deduplicate facts, reconcile contradictions, and topics.

Permission tagging

Every fact carries its own permission stamp — two people open the same page, each sees only what they're allowed to.

The injection pipeline, 8 small agents deep — every fact keeps its citation and its permission stamp from the first step.

What did we promise this account on renewal?

Answering as PM

Pricing holds for 12 months, and the team agreed not to commit to migration help before scoping it. Nothing I can see records a final answer on migration.

Answering as Sales

Pricing holds for 12 months, and migration was promised free for the first quarter — committed on the renewal call, after the team had flagged it as unscoped.

One question, one wiki. Sales is cleared for the call transcript where the promise was actually made; the PM is not, so their answer stops at its own edge instead of guessing.

Beever Atlas answering a cited question in a chat interface — Ask a plain-language question, get a cited answer streamed back — sourced to the original message.

Full model-selection table

Pull memory into an agent

1
A teammate is mid-task in Claude Code.
2
The agent calls the Beever Atlas skill over MCP.
3
It gets cited facts back, mirrored to that person's permissions.

claude code

> /beever-atlas-memory "what's our GTM decision?"

⏺ beever-atlas MCP · ask_wiki

⎿ 1 wiki page cited — Marketing & Biz

The manager decided to go ahead with the GTM strategy A at 2026-07-24.

From channel to space

The Space comes first; sources and people attach to it.

One field forced to do four jobs → one container built for two relationships.

v1 · One field

channel

Never built to do four jobs — it just ended up doing them all.

ScopePermissionsIdentitySync target

v2 · Container

Space

Built to do one job — sources and people attach to it on purpose.

Sources

Members

Beever Atlas v2 connector setup for a Space — v2 — connectors attach to a Space instead of a channel: Slack, Discord and Teams today, Drive and Jira landing next.

Why do we know if that's better?

BeeverBench: a synthetic benchmark built to answer one question: Does the agent actually retrieve information well?

71.2%
atlas-deep accuracy: +34.2pts
accuracy delta vs. rag-rerank: 75.0%
grounding precision: 46%
wiki coverage recall

BeeverBench field report — same 92-question set, four arms. atlas-quick and atlas-deep roughly double rag-base/rag-rerank on accuracy.

Full evaluation framework — 9 experiments, 3 stages

Decisions from different AI Product aspects

Even a fully AI-approved fix still needs a human to say go.

Deterministic

Leak check

Access control and the leak check — resolved against the asker's real identity, never inferred by a model.

Fact-level RBAC

Fact-level RBAC — every fact carries its own permission stamp, so the check runs per fact rather than per channel.

Gray zone

Budget gap

Clients want deterministic BI, and the early adopters who want it first don't have the budget for it.

Code-exec gateway

Apply the AI agents' code-execution logic as the injection and extraction gateway — code pulls the exact numbers, the agent only frames them.

Probabilistic

Answer variance

Extraction, synthesis and answers are LLM-driven — the same question can come back worded differently, and buyers read that as unreliability.

Published benchmark

Built BeeverBench — 92 questions across four arms — so the variance is measured and published rather than argued about.

Impact & Results

Growth highlights and areas of achievements.

#453
of 18,000+ MCP servers: 6
AI-agent discovery registries: 3
sectors show enterprise interest

Why open-core, not open-source

Open source is mainly for the fund-raising purposes. Elasticsearch open-sourced fully, AWS forked it into a free managed service, and Elastic had to retreat to a restrictive licence to survive. Features commoditise; operations don't. So the core stays MIT and the enterprise layer sells SSO, audit and compliance.

Cost and unit economics

AI workflow at a glance

I directed the design and brand work myself; the engineering team ran its own named, multi-model pipeline for anything higher-stakes.

Claude Code — Multi-agent harness (OMC) for the explainer site and my own UX/brand work — typed roles: explore, executor, reviewer, security-reviewer.
Mermaid — Architecture diagrams and internal docs, all authored as SVG/code — a model change updates the diagram, not a redraw.
Langfuse — AI observability — usage and metering records, plus LLM-as-a-judge evaluation on live traces.
GitHub — CI gate — CodeQL static analysis, coverage thresholds, typecheck and format checks, before any merge.
Linear — 745 tracked tickets, RES-### prefix, the Research team's full backlog and roadmap.

How a ticket actually moves — Mattermost and CSV into Linear, Linear into the AI harness, GitHub CI, then a human gate before production.

Teammates involved

Agents did the typing. These are the people who set the direction, argued the trade-offs and reviewed what shipped.

The research lead

CTO

AI engineer

Legal