Hippocortex/Docs
GitHubDashboard

Collective Brain

The Collective Brain is a cross-tenant knowledge layer that promotes universally useful knowledge so all agents benefit from the collective experience of the platform.

Feature flag: HIPPOCORTEX_COLLECTIVE_BRAIN (enabled by default)

What Gets Promoted

  • Problem solutions — how to fix specific errors
  • Tool usage tips — better ways to use tools and APIs
  • Efficiency techniques — faster ways to accomplish tasks
  • Error resolutions — root causes and workarounds
  • Step-by-step procedures — proven workflows

What Does NOT Get Promoted

  • Personal preferences, status updates, internal project context
  • Meeting notes, anything referencing specific people or companies
  • Content without at least 2 clear universal knowledge signals

How It Works

1. Classification

Examines compiled artifacts and assigns a verdict: universal (2+ universal signals, 0 personal signals), personal (any personal signal rejects), or ambiguous (not promoted).

2. Quality Gate

Checks specificity, actionability, confidence, and overlap with existing Brain entries.

3. Privacy Gate

Five-layer local filter (no external API calls):

  1. Category whitelist — only safe categories pass
  2. Regex PII detection — emails, phones, IPs, API keys, user paths
  3. NER heuristics — person names, organization names
  4. Semantic privacy — salary, medical, financial content
  5. Universal knowledge verification

4. Trust Scoring

Based on source count, use count, confirmations, and anomaly detection.

Brain Feedback Loop

Feature flag: HIPPOCORTEX_BRAIN_FEEDBACK_LOOP (enabled by default)

Bidirectional learning using five signals: specificity (30%), actionability (30%), recency (15%), confirmations (15%), conciseness (10%).

Four actions: agent_better (update Brain), brain_better (agent adopts), equivalent (confirm), novel (promote).

Brain Retrieval

Brain items contribute to synthesize() context, clearly attributed with [Source: Collective Brain].

Privacy Guarantees

No tenant identifiers stored. All content scrubbed. PII redacted. Organization names removed. No external API calls during privacy processing. Tenant memories remain isolated.

Configuration

HIPPOCORTEX_COLLECTIVE_BRAIN=false      # Disable entirely
HIPPOCORTEX_BRAIN_FEEDBACK_LOOP=false   # Disable feedback loop only