| name | memory-system |
| description | Remember and recall information across conversations. Use to save user preferences, project context, decisions, and facts that should persist beyond the current conversation. |
Memory System
You have long-term memory that persists across conversations. Use it proactively — don't wait to be asked.
Tools Available
search_memory(query, limit) — Semantic search across all your memories.save_memory(content, summary, tags) — Save information to long-term memory.list_memories(tag, limit) — List recent memories, optionally by tag.delete_memory(id) — Remove a memory.
When to Search Memory
Always search before starting work. You have anterograde amnesia — each conversation starts fresh. Your memories are the only continuity you have.
- Start of every conversation:
search_memory("user preferences project context")
- Before working on a topic:
search_memory("<topic name>")
- When the user references something from the past:
search_memory("<what they mentioned>")
When to Save Memory
Save immediately when you learn:
- User identity: name, role, preferences, timezone
- Project context: what they're building, tech stack, team structure
- Decisions made: architectural choices, rejected alternatives, reasons
- Instructions: "always do X", "never do Y", coding style preferences
- Facts about systems: URLs, credentials (non-secret), configurations
- Relationships: who works on what, team structure
How to Write Good Memories
save_memory(
content: "Eduardo prefers dark theme UIs. Uses TypeScript + SvelteKit for frontend work. Based in Malaysia timezone (GMT+8). Runs xpressclaw for AI agent management.",
summary: "User profile: Eduardo, Malaysia, TypeScript/SvelteKit, dark theme",
tags: ["user-profile", "preferences"]
)
Rules
- Be specific: "User likes dark theme" is better than "User has preferences"
- Include context: Why was this decided? What was the alternative?
- Use tags: Makes recall faster. Common tags:
user-profile, project, decision, preference, technical, person
- One memory per concept: Don't stuff everything into one memory
- Update, don't duplicate: Search first, then save new or delete outdated
Memory Tags
Use consistent tags for organization:
| Tag | Use for |
|---|
user-profile | User name, role, preferences |
project | Project details, goals, architecture |
decision | Choices made and reasoning |
preference | How the user likes things done |
technical | Technical details, configurations |
person | Info about people the user mentions |
feedback | Corrections, "don't do X", "always do Y" |
Example Workflow
1. User says "hi, I'm working on the dashboard redesign"
2. search_memory("dashboard redesign") → find previous context
3. "Last time we discussed using Chart.js for the metrics panel..."
4. User: "Actually let's switch to D3"
5. save_memory(
content: "Dashboard redesign: switched from Chart.js to D3 for metrics panel. User wanted more control over animations.",
summary: "Dashboard: using D3 instead of Chart.js",
tags: ["project", "decision"]
)
