Collective Cognition

Cortexa's collective cognition layer transforms your vault from a passive memory store into an active thinking system. Three interconnected features work together: background reflection analyzes your thoughts for patterns, cross-user synthesis combines perspectives from Hive peers, and confidence decay ensures stale knowledge gets flagged for review.

All three features are off by default and opt-in via the Settings > Cognition group (Reflection, Confidence, Observability, Social Memory, Procedures, Type Actions).

Background Reflection

When enabled, Cortexa periodically analyzes your recent thoughts as a collection — looking for cross-cutting patterns that you might not notice when capturing one thought at a time.

What it produces

Each reflection pass generates a type:reflection item containing:

• Themes: 3-7 recurring topics across your recent thoughts, with the specific item IDs involved.
• Contradictions: Pairs of thoughts that conflict with each other.
• Patterns: What you're spending the most cognitive energy on.
• Gaps: Unanswered questions and adjacent topics worth exploring.
• Connections: Thoughts that should be linked but aren't.
• Narrative: A 2-3 paragraph synthesis of what your recent thoughts collectively reveal.

Configuration

In Settings > Cognition > Reflection:

Setting	Default	Description
Enabled	Off	Toggle background reflection on/off
Interval (hours)	6	How often to run a reflection pass
Recent items	30	How many recent items to include in each pass

Click Run now to trigger a reflection immediately without waiting for the next scheduled interval.

How it works

1. The scheduler checks every 60 seconds whether it's time for a reflection.
2. It loads the most recent N items captured since the last reflection.
3. For each item, it reads the content, summary, and thesis core (if available).
4. Everything is sent to your configured LLM with a structured analysis prompt.
5. The LLM response is parsed for themes, contradictions, patterns, etc.
6. The output is stored as a new thought tagged type:reflection and auto:reflection.
7. If confidence scoring is enabled, the decay calculation runs during this pass.

Reflection items appear in your normal feed and are searchable. They participate in the knowledge graph just like any other thought.

Cross-User Thesis Synthesis

When you share thoughts with peers through Hive, each person develops their own thesis on a topic. Cross-user synthesis automatically combines these perspectives when related theses arrive from peers.

How it works

1. You import a thought from a Hive peer that includes thesis data.
2. Cortexa embeds the peer's content and searches your local vault for similar thoughts that also have theses.
3. If a strong match is found, an LLM synthesizes both perspectives into a cross-synthesis.
4. The output includes: agreements, disagreements, novel combinations, and a unified synthesis.
5. The result is stored as a type:cross-synthesis item with links to both source thoughts.

Configuration

Cross-user synthesis is controlled by a single toggle in Settings > Cognition > Reflection:

Setting	Default	Description
Cross-user thesis synthesis	Off	Enable automatic synthesis when importing peer theses

This feature requires both reflection to be enabled and at least one Hive connection with auto-import active.

Memory Confidence & Decay

Knowledge gets stale. A decision you made 6 months ago might no longer apply. Confidence scoring adds a numerical quality signal to every thought that naturally decays over time, flagging old, unreviewed knowledge for attention.

How scoring works

Every thought gets an initial confidence score based on its source:

Source	Initial Score
Manual capture	0.80
MCP / API / CLI	0.70
Hive shared	0.60
Telegram	0.50

How decay works

Each week, every thought's confidence decays by a configurable rate:

new_score = current_score × (decay_rate ^ weeks_since_last_validated)

With the default rate of 95% per week:

• After 5 weeks: ~77% of original score
• After 10 weeks: ~60% of original score
• After ~23 weeks: approaches the stale threshold

Stale tagging

Thoughts that fall below the stale threshold (default 30%) are automatically tagged confidence:stale. You can filter by this tag to review aging knowledge. Items above the threshold have the tag removed automatically.

Resetting decay

User interactions that show engagement with a thought reset the decay clock:

• Pinning a thought resets last_validated_at to now.

This means thoughts you actively reference stay confident, while forgotten ones naturally age.

Contradiction penalties

When the reflection pass detects contradicting thoughts, both items receive a confidence penalty (default 20% reduction). This surfaces conflicting knowledge faster.

Configuration

In Settings > Cognition > Confidence:

Setting	Default	Description
Enabled	Off	Toggle confidence scoring on/off
Decay rate (per week)	95%	Higher = slower decay
Stale threshold	30%	Items below this are tagged confidence:stale
Contradiction penalty	20%	Score reduction when contradictions are found

Visual indicators

When confidence is enabled, item cards show a colored badge:

• confident (green): score >= 70%
• aging (yellow): score 40-70%
• stale (red): score < 40%

How the three features interact

The three features form a reinforcing loop:

1. You capture thoughts throughout the week.
2. Reflection runs every N hours, analyzing patterns and detecting contradictions.
3. Confidence decay runs during each reflection pass, aging unreviewed knowledge.
4. When a Hive peer shares a thought with a thesis, cross-synthesis finds your related thoughts and produces a combined view.
5. The synthesis and reflection outputs are themselves thoughts — they get embedded, searched, and eventually reflected upon in future passes.

Over time, your vault develops a meta-layer of self-awareness: it knows what themes dominate your thinking, where your knowledge conflicts, what's getting stale, and how your perspectives compare to your peers.

Tool Observability

Every LLM call, embedding operation, and orchestration step is logged to a persistent tool_invocations table. This gives you visibility into what Cortexa is doing behind the scenes.

What it tracks

• Tool name and category (llm, embed, search, tts)
• Success/failure status
• Duration in milliseconds
• Token usage (for LLM calls)
• Associated item and job IDs

Configuration

In Settings > Cognition > Observability:

Setting	Default	Description
Enabled	Off	Toggle tool invocation logging
Retention (days)	30	Auto-purge invocations older than this

Social Memory

When you connect to Hive peers, Cortexa tracks their expertise and contribution patterns over time.

What it tracks

Per-peer profiles auto-update on every import:

• Item count — how many thoughts this peer has shared
• Thesis count — how many of their shares included thesis analysis
• Average thesis confidence — quality signal for their contributions
• Trust score — synthesized credibility (starts at 0.5, increases with quality contributions)

Configuration

In Settings > Cognition > Social Memory:

Setting	Default	Description
Enabled	Off	Toggle peer tracking
Track expertise	Off	Derive expertise from thesis quality
Initial trust score	50%	Starting trust for new peers

Procedural Memory

Cortexa learns from how the orchestration engine solves problems. Every orchestrator execution logs its tool sequence and outcome. Over time, successful patterns emerge.

How it works

1. The orchestrator runs a tool sequence for a thought (e.g., decompose → challenge_claims → synthesize_brief).
2. The sequence and outcome are logged to procedure_log.
3. Patterns that repeat successfully are promoted to learned_procedures with confidence scores.
4. When "suggest procedures" is enabled, the orchestrator is given these patterns as hints for future planning.

Configuration

In Settings > Cognition > Procedures:

Setting	Default	Description
Enabled	Off	Toggle procedural memory
Suggest procedures	Off	Feed learned patterns to the orchestrator
Min executions to learn	3	How many times a pattern must succeed before it's learned

Thought Type Actions

Configure which auto-actions run when agent captures create items of each thought type. This lets you customize the pipeline per type — for example, always run Deep Think on architecture decisions, but only summarize notes.

Configuration

In Settings > Cognition > Type Actions:

Each of the 8 thought types (decision, discovery, architecture, debug, pattern, question, todo, note) has 4 toggle chips:

• Summarize — auto-generate a summary
• Deep Think — run the full thesis pipeline
• Research — trigger web + vault research
• Brief — generate an audio briefing

Defaults

Type	Summarize	Deep Think	Research	Brief
decision	Yes	No	No	No
discovery	Yes	No	No	No
architecture	Yes	No	No	No
debug	Yes	No	No	No
pattern	Yes	No	No	No
question	No	No	Yes	No
todo	No	No	No	No
note	No	No	No	No

Memory Governance (Search Boosting)

The cognition stack feeds back into search ranking. When confidence scoring is enabled, FTS search results are boosted by confidence:

boosted_score = raw_bm25_score × (0.7 + 0.3 × confidence_score)

This means high-confidence items (recently validated, well-researched theses) rank higher than stale or low-confidence items for the same query. The effect is subtle but compounds — your best knowledge surfaces first.

The Full Picture

All 7 layers form an interconnected system:

  Capture → Embed → Store
       ↓
  Reflection → finds themes, contradictions, patterns
       ↓
  Confidence → decays stale items, boosts well-researched
       ↓
  Cross-Synthesis → merges peer perspectives
       ↓
  Observability → tracks what works
       ↓
  Procedural Memory → learns successful workflows
       ↓
  Social Memory → tracks peer expertise
       ↓
  Memory Governance → search ranking reflects all signals
       ↓
  Agents get the best context automatically

Over time, your vault develops a meta-layer of self-awareness: it knows what themes dominate your thinking, where your knowledge conflicts, what's getting stale, how your perspectives compare to your peers, and which analysis approaches work best for different types of problems.