Graphs & Sensitivity Tiers

Graphs let you organize your Cortex into distinct namespaces. Sensitivity tiers let you decide how much of each graph an AI is allowed to see — including the option to block AI access entirely.

Important: even for graphs with open tiers, your AI never receives everything in that graph. On every recall call, Graphnosis performs a semantic search and returns only the nodes most relevant to your current question, subject to the token cap for that graph. The rest stays encrypted and untouched. This is intentional — Graphnosis acts as your hippocampus, retrieving targeted memory traces, not dumping your files.

Multiple graphs

A single Cortex can hold multiple graphs. You might create:

• work — project notes, meeting summaries, technical decisions
• personal — journal entries, health notes, finance snippets
• research — papers, articles, reference material

Each graph is independent: its own sensitivity tier, its own token budget, its own sources. When the AI calls recall, it searches across all graphs you’ve granted it access to, respecting each graph’s tier.

Creating a graph

Open the Graphnosis window → click + New Graph → give it a name and choose a sensitivity tier.

Sensitivity tiers

Every graph is assigned one of three tiers:

Tier	What it means
public	Content may be surfaced to any AI client without restriction.
personal	Content is surfaced only when the AI explicitly asks for context (no proactive injection). Token cap applies.
sensitive	Content is never surfaced to AI clients automatically. Requires manual export or explicit user action.

Tiers are a hard cap, enforced by the sidecar before any content leaves the Cortex. The AI model itself never sees the tier configuration — it simply doesn’t receive content above its allowed level.

Tier behavior in practice

public graphs — chunks flow freely into AI context. Proactive injection is enabled. Best for reference material, documentation, public notes.

personal graphs — chunks are only returned when recall is called explicitly. Proactive injection is disabled. The token cap limits how much context can be returned per conversation turn. Best for personal notes, journal entries, work summaries.

sensitive graphs — the sidecar returns zero results for recall queries targeting this graph. The AI is never told why — it just gets no results. You can still search and review these memories in the Graphnosis UI. Best for health information, financial records, anything you want to keep entirely local.

Per-graph token caps

Each graph has a configurable maximum token budget per recall response. This prevents any single graph from consuming the entire context window of your AI.

Defaults:

Tier	Default token cap
public	4000 tokens
personal	2000 tokens
sensitive	0 (AI access blocked)

You can override these in policy.json.

Configuring policy.json

policy.json lives in the root of your Cortex folder. It is not encrypted (it contains only policy rules, not content). You can edit it directly:

{
  "graphs": {
    "work": {
      "tier": "personal",
      "maxTokensPerRecall": 3000
    },
    "research": {
      "tier": "public",
      "maxTokensPerRecall": 6000
    },
    "health": {
      "tier": "sensitive"
    }
  },
  "globalMaxTokensPerTurn": 8000
}

Changes to policy.json take effect immediately — no restart required. The sidecar watches the file for changes.

globalMaxTokensPerTurn

This is a hard ceiling across all graphs combined. Even if individual graph caps add up to more, the sidecar will truncate the total context to this limit. Default: 8000.

Which graphs does a client see?

By default, all graphs in the Cortex are visible to any connected MCP client, subject to their tier. If you want to expose only specific graphs to a specific AI client, you can scope the sidecar using the GRAPHNOSIS_GRAPHS environment variable (see Environment Variables).