NewEarly access is open. The first 20 teams get a locked rate for 12 months.
RELAY
Sign inBook a demo

Knowledge infrastructure for AI-native teams

Organizational memory,
atom by atom.

Your AI assistant doesn't know what your team decided last Tuesday. Relay captures every commitment, decision, and policy from meetings, threads, and tickets, and feeds them to the tools you already use.

Try it live →

SOC 2 Type 2 · Your data. Zero training.

ATOM · LIVE EXAMPLE

DECISION
a-2841

Refund window locked at 30 days for standard, 60 days for enterprise.

SPEAKER
Sarah Chen
SOURCE
Refund discussion · 2026-05-01
WHEN
2026-05-01
CONFIDENCE
0.94
VERSION
v3 · supersedes a-2719
CITE
[a-2841]
Type your own

EXTRACTING · LAST 60S

STREAM
  • DECISION

    Refund window locked at 30 days standard, 60 enterprise.

    Meet · Refund discussion

    [a-2841]
  • COMMITMENT

    Sarah owns escalations over $5k.

    Slack · #ops

    [a-2902]
  • RISK

    Postgres 16 upgrade blocked on RLS policy migration.

    Linear · ENG-1184

    [a-2910]

0+

Atoms processed

0+

Citations rendered

0

Conflicts surfaced

0.0s

Median atom latency

Sources

Reads where the work happens.

Adapters connect over the existing OAuth / webhook surface. No agent installed on endpoints.

The primitive

An atom is one unit of organizational truth.

Every claim is one sentence: said by someone, in a source you can open, on a date you can check. Scroll to see what's inside.

ATOM · ANATOMY

COMMITMENT
a-2902

Sarah owns escalations over $5k. Looped in legal for contract exceptions.

SPEAKER
Marcus Lee
SOURCE
Ops sync · 2026-05-08
WHEN
2026-05-08
CONFIDENCE
0.91
VERSION
v1
CITE
[a-2902]
  1. Type chip01 / 06

    DECISION · COMMITMENT · RISK · FACT · QUESTION · BLOCKER

    A closed set. Drives downstream routing: commitments fire reminders, risks feed drift detection.

  2. Claim02 / 06

    One sentence. One claim.

    If a claim has an 'and', the extractor probably should have split it.

  3. Provenance03 / 06

    speaker · source · timestamp

    Every atom traces back to a transcript line, a Slack permalink, or a doc anchor.

  4. Confidence04 / 06

    0.91

    LLM-graded, validated against a 4-tier classifier (auto · surface · mirror · drop).

  5. Version05 / 06

    v1 · supersedes a-NNNN

    When the org changes its mind, the new atom supersedes the old. Old atoms are never deleted.

  6. Citation ID06 / 06

    [a-2902]

    Stable across renderings. The same ID surfaces in Slack, in Claude, in the dashboard.

Try it

Ask your org a question.

Type anything, or pick an example. Relay returns a one-line answer backed by the exact atoms it came from. These are canned examples — no live model — but the shape is exactly what the API returns.

Try

Ask a question to see a cited answer. Canned examples — no live model.

Everything in one substrate

One memory layer. Six capabilities your AI can call.

Atoms are the primitive. Everything else is what you do with them. Each capability is a deliberate decision documented in the architecture spec.

Atomic extraction

Every claim, isolated and cited.

A transcript becomes 14 atoms. Each one carries a speaker, a source, a timestamp, a confidence, and a stable cite ID, usable anywhere.

[a-2841]→[a-2902]→[a-2945]… 11 more

Versioning

Decisions supersede. Never delete.

When the org changes its mind, the new atom replaces the old; the chain stays inspectable.

Conflict detection

Two atoms disagree? We say so.

Embedding similarity flags contradictions; a steward applies the supersedence with one click.

Knowledge graph

Atoms, entities, conflicts, all queryable.

Internal context for synthesize-query (GraphRAG) and an explorable surface for stewards.

SDKs

Python · TypeScript · MCP.

Plus renderers for LangChain, Semantic Kernel, Claude Projects.

Trust ladder

Mirror → Autonomous, one step at a time.

Promote per pipeline. Step back down automatically if rejections spike.

The pipeline

From raw conversation to cited answer.

Five stages, every one logged. Deterministic where it can be, LLM-bounded where it can't. Scroll to travel through them.

STAGE 018 adapters

Sources

slack · meet · github · notion · gmail · linear

OAuth + webhook adapters pull from where the work actually happens. No endpoint agent installed, no daemon to update.

SCROLL
STAGE 02LLM-bounded

Extract

claim · speaker · timestamp

A pinned LLM call splits transcripts and threads into single-claim atoms with provenance attached. Every prompt is logged.

NEXT
STAGE 03deterministic

Resolve

entities · domains · ownership

Entities link to the people graph; embedding clusters assign each atom to a domain (refunds, hiring, infra…). Replayable end-to-end.

NEXT
STAGE 04audit-logged

Version

supersede · merge · audit

New atoms supersede old ones; conflicts open as first-class records. Nothing is ever silently deleted. The version chain is the truth.

NEXT
STAGE 05API · SDK · MCP

Atoms

indexed · cited · served

Atoms ship to the dashboard, the Python + TypeScript SDKs, and any tool that speaks MCP or HTTP. One memory, many surfaces.

CONTINUE

Swipe →

Conflict resolution

When the org contradicts itself.

Relay detects when two claims disagree, scores their similarity, and proposes a supersedence. You keep the final say. Pick a resolution to preview.

POLICY CONFLICTc-104 · 92% sim

Two policies disagree on enterprise refund window.

DECISION
a-2841

Enterprise customers get 60 days for refunds.

SPEAKER
Sarah Chen
SOURCE
Refund discussion · 2026-05-01
WHEN
2026-05-01
CONFIDENCE
0.94
VERSION
v3
CITE
[a-2841]
DECISION
a-1980

All customers, including enterprise, capped at 30 days.

SPEAKER
Marcus Lee
SOURCE
Pricing review · 2026-04-12
WHEN
2026-04-12
CONFIDENCE
0.88
VERSION
v1
CITE
[a-1980]

Relay suggests · Keep A · supersede B

The May 1 decision supersedes the April 12 ruling for enterprise tier. Keep [a-2841] active and mark [a-1980] superseded for the enterprise segment. Standard tier stays at 30 days.

The graph

Every atom is a node.

People, atoms, and conflicts connect into one map of what your company knows. Scroll to watch it assemble.

atomentityconflict
Sarah ChenMarcus LeeRefund discussionPricing reviewLegal · $5k capa-2841a-1980a-2902a-2750a-2799a-2814c-104

12 nodes · 14 edges · 1 conflict

Positioning

Why not just search your docs?

Glean, Notion AI, Otter: every adjacent product treats a document or a transcript as the unit of knowledge. Relay treats the claim as the unit.

CapabilityRelayGlean / Notion AIOtter / Fireflies
Knowledge unitAtomic claims (cited)Full documentsTranscript chunks
VersioningSupersedence chain, audit logNot modeledNot modeled
Conflict detectionFirst-class model + steward UINot modeledNot modeled
ProvenancePer-claim speaker + source + tsPer-documentPer-transcript
Agent distributionMCP · LangChain · SK · ClaudeSearch APINot modeled
Trust ladder (autonomy gating)Mirror → Autonomous, per pipelineNot modeledNot modeled

Trust ladder

You decide how far Relay reaches.

Every org starts in watch-only mode. Promote one step at a time, with full audit trails. If Relay gets things wrong too often, it steps itself back down.

MirrorSuggestionLight actionAutonomous

STAGE 1

Mirror

DEFAULT

Watch and extract.

  • Atoms surface in your workspace
  • Skills compile as drafts
  • Zero external writes; default for every new org

STAGE 2

Suggestion

Propose, never push.

  • High-confidence atoms surface in Brief
  • Skills enter steward review
  • Distribution only for approved skills

STAGE 3

Light action

Publish skills. Send nudges.

  • Auto-publish on confidence threshold
  • Internal nudges (DMs, reminders) allowed
  • Still zero external write-back

STAGE 4

Autonomous

Act, with citations.

  • Relay can write to Linear, Slack, and Notion, only to lists you approve.
  • Every action cites the atom it acted on.
  • Steps itself down a level if it gets things wrong.

Relay steps down a level automatically if its suggestions get rejected too often, if an action it took had to be reverted, or if a privacy rule was tripped.

Distribution

Plugs into your stack.

One read-only API. Official Python + TypeScript SDKs. A native MCP server. Drop-in plugins for LangChain, Semantic Kernel, and Claude Projects.

pip install relay-skills
from relay_skills import Relay

relay = Relay(org="acme", api_key=os.environ["RELAY_KEY"])

# Today's brief: all atoms changed since last fetch
brief = relay.brief.today()
for atom in brief.decisions:
    print(atom.claim, atom.cite_id)

# Ask Relay a question; every fact comes cited
answer = relay.query("What's our refund window?")
print(answer.text)               # → "30 days standard, 60 enterprise [a-2841]"
print(answer.sources)            # → [{cite_id: 'a-2841', confidence: 0.94}]
MCP server

MCP server

stdio · http

LangChain

LangChain

renderer

Semantic Kernel

Semantic Kernel

renderer

Claude Project

Claude Project

bundle

Works with

  • AnthropicAnthropic
  • ClaudeClaude
  • LangChainLangChain
  • Hugging FaceHugging Face
  • VercelVercel

Pricing

One tier. Locked rate for 12 months.

We're onboarding the first 20 teams. Direct line to the founders, weekly office hours, and a permanent locked-rate guarantee.

Standard

For the first 20 teams to onboard.

AVAILABLE
  • Unlimited atoms, skills, and citations
  • All adapters: Slack, Meet, GitHub, Linear, Notion, Gmail, Teams, Zoom
  • Full trust ladder: Mirror → Suggestion → Light action → Autonomous
  • Python + TypeScript SDKs, MCP server, LangChain + SK renderers
  • Rate locked for 12 months from kickoff. No renegotiation at scale
  • Direct line to engineering · weekly office hours
Contact us

Questions, answered

Scope, data, exit.

The five questions every prospective customer has asked. If yours isn't here, we're one email away.

All adapters, both SDKs, the MCP server, the full trust ladder, and a direct line to engineering. Rate is locked for 12 months from kickoff. No renegotiation when you scale headcount inside that window.

Every row in Postgres is scoped by tenant via Row-Level Security; the adapter layer never sees cross-tenant data; the LLM client always flushes credentials between calls. Defense in depth is documented in the architecture spec.

No. Your atoms, transcripts, and source documents are never used to train any model, ours or any third party's. The LLM API providers we use (OpenAI, Anthropic) operate under zero-retention agreements for Relay traffic.

One read-only HTTP API, official Python and TypeScript SDKs, an MCP server, and renderers for LangChain, Semantic Kernel, and Claude Projects. No webhooks to chase, no custom protocol to learn. Code samples are in the Distribution section above.

Export every atom, every citation, and every conflict as JSON or Markdown, including the full version chain. Skill bundles export as plain SKILL.md files that work with any MCP-aware tool, with or without Relay.