Research & Memory

Claude Code includes a built-in deep research system for source-grounded investigations and a semantic memory layer for cross-session knowledge retention.

Deep Research

The research system (src/research/) enables thorough, source-grounded investigations across local files, wiki content, and memory.

Capabilities

Feature	Description
Citation Extraction	Automatically extracts and attributes sources from research results
Claim Verification	Cross-references claims across sources with confidence scoring
Dossier Generation	Produces structured research dossiers with summaries, evidence, and references
Truth Checking	Validates factual claims against source material
Source Ranking	Ranks sources by relevance, authority, and recency

Research Workflow

Query — Natural language research question or topic
Gather — Collects information from local files, wiki, memory, and web
Analyze — Cross-references, verifies claims, extracts citations
Synthesize — Produces a structured report with evidence and sources

/research "What are the current performance bottlenecks in the API layer?"

  # Result: structured dossier with citations, evidence, and ranked sources

Semantic Memory

The memory system (src/memdir/) provides persistent, cross-session knowledge retention using text embedding search.

Features

Feature	Description
Semantic Search	Find memories by meaning, not just keywords — cross-lingual support
Auto-Memory Capture	Important context is automatically saved during conversation
Memory Age Tracking	Older memories decay in relevance; fresh memories are prioritized
Hybrid Recall	Combines local embedding search with LLM-assisted retrieval
Team Sync	Share memories across sessions and team members

Memory Commands

/memory                          # Edit Claude memory files
  /memory-search "deployment config"  # Semantic search across memories
  /research "auth patterns"           # Research across files + wiki + memory

How Memory Works

Memories are stored as markdown files in the user's Claude memory directory. When a query is made:

Text embeddings are computed for the query
Embeddings are compared against stored memory vectors
Top matches are ranked by similarity and recency
Matched memories are injected as context for the LLM

Note: Memory is distinct from session history. Session history records raw conversation turns; memory captures curated facts, preferences, and project context for long-term retention.

Architecture Files

Path	Role
`src/research/`	Deep research: citation extraction, dossier generation, truth checking, source ranking
`src/memdir/`	Semantic memory: text embedding search, memory age tracking, cross-session recall