set-core includes a persistent memory system (shodh-memory) that gives agents cross-session recall — decisions, learnings, and context survive across conversations and are shared between agents.
Memory operates through hooks that fire automatically during Claude Code sessions:
| Hook | When | What |
|---|---|---|
| Warmstart | Session start | Loads relevant memories as context |
| Pre-tool | Before each tool call | Injects topic-based recall |
| Post-tool | After Read/Bash | Surfaces past experience for the file/command |
| Save | Session end | Extracts and saves new insights |
Agents don't need to explicitly save — the infrastructure handles it.
set-memory health # check memory system status
set-memory stats # memory statistics
set-memory recall "query" # semantic search
set-memory list # list all memories
set-memory forget <id> # remove a memory
The memory page shows health status, type breakdown, and retrieval statistics:
The memory system is also available as MCP tools for programmatic access:
| Tool | Purpose |
|---|---|
remember |
Save a memory with type and tags |
recall |
Semantic search |
brain |
3-tier memory visualization |
context_summary |
Condensed summary by category |
proactive_context |
Topic-aware context injection |
forget |
Remove a memory |
list_memories |
List with filters |
memory_stats |
Health and metrics |
pip install 'shodh-memory>=0.1.81'
set-memory health # verifyMemory degrades gracefully — if not installed, all memory commands silently no-op. Agents work fine without it, they just don't have cross-session recall.
Memories are categorized by type, which affects retrieval priority and display:
| Type | Purpose | Example |
|---|---|---|
| Decision | Architectural or design choices | "Chose eager merge policy for this project" |
| Learning | Patterns discovered during development | "vitest needs --passWithNoTests flag" |
| Context | Project state and configuration | "Project uses Prisma with PostgreSQL" |
| Bug | Known issues and their fixes | "Playwright flaky on CI without --retries=2" |
| Feedback | User preferences and corrections | "Never modify consumer project code" |
Most memories are saved automatically by the session-end hook. For important insights you want to ensure are captured with high priority:
echo "Always use --project-type web for web projects" | set-memory remember --type Decision --tags source:user,initTo correct a wrong memory or suppress a false positive:
set-memory forget <memory-id>Memories are shared across all agents in the same project. When agent A discovers that a test requires a specific flag, agent B will receive that context in its next session through the warmstart hook. This eliminates redundant investigation across parallel worktrees.
Agents do not save memories voluntarily -- across 15+ sessions in benchmarks, zero voluntary saves were observed. The hook-driven infrastructure is essential for building useful cross-session recall.