Concepts

How Relay organizes, scores, and delivers project context across every AI tool you use.

Projects

A project is the top-level container in Relay. Each project has its own context, memory, packets, and work sessions. Think of it as one codebase, one initiative, or one product.

What is Project context?

Project Context is your project's stable truth — the facts Relay is confident about and uses to generate briefs. Context entries are grouped by kind:

objective

What the project is trying to achieve right now.

decision

A choice that was made and should be respected.

constraint

A rule or limitation that must hold.

task

An open work item.

progress

Recent accomplishments or milestones.

architecture_fact

Structural knowledge about the codebase.

risk

Known risks or potential issues.

assumption

Something taken as true without proof.

Each entry has a status: active (confirmed), tentative (low confidence, pending review), disputed (conflicting evidence), superseded (replaced by a newer entry), stale (no recent reaffirmation), or resolved (completed/closed).

You can lock any entry to prevent Relay from modifying it automatically.

How Relay decides what is true

Relay builds context through an observe → reflect → promote cycle:

  1. Observe — Each session digest is analyzed for facts that might be context-worthy.
  2. Reflect — Observed facts are compared against existing context for conflicts or confirmation.
  3. Promote or conflict — Strong evidence promotes tentative entries to active; contradictions flag disputes.

Confidence scores, evidence count, and reaffirmation frequency determine promotion speed. On aggressive autonomy, Relay promotes eagerly. On conservative, everything stays tentative until you lock it.

Current vs historical truth

Context entries have optional validFrom and validUntil timestamps. Active entries represent current truth; superseded entries represent what was true before.

When you ask Relay “what was the objective last week?”, temporal retrieval checks historical context and superseded entries. Current queries use active context only.

Why memory may be demoted or archived

Raw memory items go through compaction. When a fact graduates into context or gets absorbed by a summary snapshot, the original memory item is demoted to save retrieval budget.

  • covered_by_context — The fact is now a context entry.
  • covered_by_summary — A project summary covers the same ground.
  • historical_only — Retained for provenance, not included in active retrieval.
  • completed — Task or note is done.

Demoted items are never deleted — they remain available for historical queries and provenance tracking. Compaction aggressiveness is controlled in project settings (light / standard / aggressive).

Browser packets vs agent packets

Context packets are the formatted output Relay injects into your AI tools. There are two families:

Fresh chat bootstrap

Full project context for a new browser chat session. Includes context, recent memory, and project state.

Quick continuity

Delta packet for continuing an existing browser session. Only what changed since last sync.

Agent full bootstrap

Deep context for MCP-connected agents (Claude Code, Cursor, etc.). Richer format with architecture facts and tool references.

Agent quick continuity

Lightweight delta for agents mid-session. Minimal token cost.

Packets are target-aware: the same context produces different output for Claude Code vs. ChatGPT based on token budgets and formatting conventions.

Memory items

Memory items are the raw captured context from your sessions. Each has a type:

decision

A choice that was made and should be remembered.

task

Something to be done or tracked.

constraint

A rule or limitation.

note

General context or observations.

requirement

A user or product requirement.

artifact

A reference to a file, URL, or resource.

Items are scored by decay (freshness + reaffirmation) and can be pinned for priority inclusion in packets. Strong memory items graduate into context entries over time.

Drift reconciliation

When you work across multiple AI tools simultaneously, each surface may produce different claims about the same topic. Relay's drift reconciler detects when facts from different surfaces contradict each other and flags them as disputes.

By default, the most recent surface “wins” for active status, but both perspectives are retained for review.

Need pricing details?

See Plans & limits for the full Free / Starter / Pro comparison.