Cortex

Returns the user's cognitive profile for the current domain. Pre-computed, <50ms. Use at session start for context injection.

Lightweight domain classification from cwd, project, or first message. <20ms.

Full rescan of all session data to rebuild methodology profiles. <10s.

Overview of all detected cognitive domains. <10ms.

Incremental profile update after a session ends. <200ms.

Returns methodology map as graph data for 3D visualization. <100ms.

Launch the unified 3D neural graph in the browser. Combines methodology profiles, memories, and knowledge graph.

Explore interpretability features: dictionary features, attribution graphs, persona vectors, and cross-domain behavioral persistence. Inspired by Anthropic's mechanistic interpretability research.

Drive the ai-architect pipeline end-to-end: discovery -> impact -> strategy -> PRD -> verification -> implementation -> PR. Connects to ai-architect MCP server over stdio.

Store a memory through the predictive coding write gate.

Retrieve memories using intent-adaptive PG recall with production enrichments.

Memory system diagnostics: counts, heat distribution, entities, triggers.

Hippocampal replay: save/restore working state across context compaction events.

Generate a project narrative/story from stored memories for a directory or domain.

Run memory maintenance: heat decay, compression, and CLS consolidation cycles.

Import conversation history from ~/.claude/projects/ into the memory store. Extracts decisions, errors, architecture notes, and key insights from JSONL sessions.

Delete a memory by ID. Supports soft (mark stale) or hard (permanent) deletion. Protected memories require force=True.

Validate memories against current filesystem state. Marks stale memories whose referenced files no longer exist.

Rate a memory as useful or not. Drives metamemory confidence which affects decay resistance and recall ranking.

Bootstrap memory from an existing codebase. Analyzes structure, config, docs, entry points, and CI/CD. Stores key discoveries as memories.

Mark a memory as compaction-resistant. Anchored memories survive context compaction, heat decay, and cannot be deleted without force. Use for critical facts and active decisions.

Auto-import prior Claude Code conversations into the memory store. Idempotent -- tracks already-processed session files by hash. Links historical work to core concepts automatically.

Analyze a codebase and store its structure as Cortex memories. Uses tree-sitter AST for cross-file resolution, call graphs, and community detection. Incremental: only processes changed files.

Retrieve memories using the fractal hierarchy (L0/L1/L2 clusters). Adaptive weighting based on query length — short queries search broad, long queries search specific.

Navigate into a fractal memory cluster. L2 cluster → shows L1 child clusters. L1 cluster → shows individual memories.

Navigate memory space using Successor Representation (co-access patterns). Starting from a memory ID, returns co-accessed neighbors and their SR distances.

Trace entity relationships through the knowledge graph. Returns causal/dependency chains from a starting entity or memory.

Identify knowledge gaps in the memory store: isolated entities, sparse domains, temporal drift, and low-heat topic clusters.

Sync top memory insights into CLAUDE.md for the project directory. Adds or refreshes a '## Memory Insights' section.

Create a prospective memory trigger: a future-oriented reminder that fires when a condition is met (keyword, time, file, or domain).

Add a neuro-symbolic rule to the memory store. Rules hard-filter or soft-rerank recall results based on conditions.

List active neuro-symbolic rules in the memory store, optionally filtered by scope or rule type.

Generate a period-based autobiographical narrative of project activity. Returns chronological 'chapters' of what happened during a time period.

Evaluate knowledge coverage completeness for a project directory. Returns a 0-100 coverage score and actionable recommendations.

Author a wiki page (adr/specs/files/notes) or append/replace an existing one. Pages live under ~/.claude/methodology/wiki/ and are indexed in PostgreSQL as protected pointer memories for recall.

Read the raw markdown of a wiki page by its relative path.

List authored wiki pages. Optionally filter by kind (adr/specs/files/notes).

Add a bidirectional link between two wiki pages. Updates the Related section of both files. Idempotent.

Create a numbered ADR (architecture decision record) from structured fields. Auto-increments the ADR number.

Regenerate the wiki table of contents at .generated/INDEX.md. Authored pages are never touched.