Knowledge Catalog
The Knowledge Catalog (Phase 5) exposes Claude Code’s local knowledge layer through 3 MCP tools. It makes project memory and prompt history searchable and accessible from any MCP client.
Data Sources
| Source | Location | Description |
|---|---|---|
| Memory files | ~/.claude/projects/{project}/memory/*.md | Curated project knowledge — the highest-signal content per project |
| History | ~/.claude/history.jsonl | Every user prompt across all sessions, with timestamps and session IDs |
MCP Tools
Memory
| Tool | Parameters | Description |
|---|---|---|
list_memory_files | project (optional filter) | List all memory files with metadata (size, modified date). No content returned — use get_memory_file for content |
get_memory_file | project (required), filename (default: MEMORY.md) | Read a specific memory file’s content |
Example queries:
- “What memory files exist across my projects?”
- “Show me the MEMORY.md for the antoncore project”
- “Read the cloudflare-ai-bots.md memory from antoncore”
History
| Tool | Parameters | Description |
|---|---|---|
search_history | query, project, session_id, limit, offset, since | Search the global prompt history with keyword, time, project, and session filters |
Example queries:
- “Find prompts where I mentioned docker”
- “Show my prompts from the last 24 hours”
- “What did I ask in session abc123?”
Architecture
Collector Integration
The background collector runs 2 scan passes for the knowledge catalog after transcript indexing:
collect_once()
├── Transcript scan (existing)
├── Memory scan → scan_memory_files() → upsert to Redis
└── History scan → parse_history() → append to Redis (incremental)
Incremental History Parsing
The history.jsonl file can grow large (1MB+). The collector tracks a byte offset and only reads new lines on each cycle:
- Seek to last known offset
- Detect line boundary (peek at byte before offset)
- Skip partial line if mid-line
- Parse new complete JSON lines
- Store updated offset
Redis Key Schema
agentibridge:sb:memory:{project}:{filename} # Hash: metadata + content
agentibridge:sb:idx:memory # Sorted set: all memory keys (score = mtime)
agentibridge:sb:history # List: JSON history entries
agentibridge:sb:pos:history # String: byte offset
File Fallback
When Redis is unavailable, all operations fall back to direct filesystem reads:
- Memory: Scans
~/.claude/projects/*/memory/*.mddirectly - History: Reads
~/.claude/history.jsonlfrom the beginning
Configuration
| Variable | Default | Description |
|---|---|---|
CLAUDE_CODE_HOME_DIR | ~/.claude | Claude Code home directory. All paths derive from this: {home}/projects/, {home}/history.jsonl, {home}/projects/{project}/memory/ |
AGENTIBRIDGE_MAX_HISTORY_ENTRIES | 5000 | Max history entries in Redis |
AGENTIBRIDGE_MAX_MEMORY_CONTENT | 51200 | Max bytes per memory file (50KB) |
Docker
The docker-compose.yml mounts the entire ~/.claude directory read-only:
volumes:
- ${CLAUDE_DIR:-~/.claude}:/home/appuser/.claude:ro
This gives the container access to transcripts, history, and memory files without separate mount entries.
See Also
- Internal Architecture — Full module reference and design patterns
- Configuration Reference — All environment variables