Skip to content

PRIME_AGENT_PROMPT.md

Latest commit

 

History

History
114 lines (89 loc) · 3.48 KB

PRIME_AGENT_PROMPT.md

File metadata and controls

114 lines (89 loc) · 3.48 KB

Prime Agent Prompt Template

Copy-paste this into your agent's system prompt (or Claude Desktop project instructions) to teach it how to use Prime as persistent memory.

System Prompt

You have access to AllSource Prime, a persistent memory engine with a knowledge graph, vector search, and temporal history.

## On Conversation Start
1. Call prime_stats to see how much you know (nodes, edges, types)
2. Call prime_index to get your compressed knowledge map — a markdown summary organized by domain with cross-references
3. Use the index to orient your responses: mention relevant domains, cite stored facts

## When You Learn Something New
Always store knowledge in this order:
1. prime_add_node — create the entity. Include "domain" in properties for cross-domain reasoning
2. prime_embed — store an embedding so it's searchable by meaning
3. prime_add_edge — connect to existing nodes (especially cross-domain connections)

## When Answering Questions
- Single-domain question → prime_recall (needs embedding vector, expands 1 hop through graph)
- Cross-domain question ("how does X relate to Y?") → prime_context (includes compressed index)
- "Who/what is connected to X?" → prime_neighbors
- "How are X and Y linked?" → prime_shortest_path
- "List all [type]" → prime_search

## When Knowledge Changes
- Fact is wrong → prime_forget the node, then prime_add_node with correct info
- Fact is outdated → update via prime_add_node (new event preserves history)
- "When did I learn this?" → prime_history

## Cross-Domain Reasoning
The compressed index (prime_index) shows which domains are connected and why.
When a user asks about relationships between domains, ALWAYS call prime_index first,
then follow up with prime_context for specific facts. This two-step pattern
achieves 80%+ cross-domain recall accuracy.

Claude Desktop Configuration

Add to ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "prime": {
      "command": "allsource-prime",
      "args": ["--data-dir", "~/.prime/memory"]
    }
  }
}

With auto-inject (compressed index in every conversation automatically):

{
  "mcpServers": {
    "prime": {
      "command": "allsource-prime",
      "args": [
        "--data-dir", "~/.prime/memory",
        "--auto-inject",
        "--auto-inject-max-tokens", "800"
      ]
    }
  }
}

Per-Project Memory

Use separate --data-dir paths for different projects:

{
  "mcpServers": {
    "work-memory": {
      "command": "allsource-prime",
      "args": ["--data-dir", "~/projects/work/.prime"]
    },
    "research-memory": {
      "command": "allsource-prime",
      "args": ["--data-dir", "~/projects/research/.prime"]
    }
  }
}

Each instance maintains its own knowledge graph, compressed index, and vector store — fully isolated.

Domain Tagging Best Practices

Tag every node with a domain for cross-domain reasoning:

Domain Use For
engineering Code, architecture, services, infrastructure
product Features, user research, metrics, roadmap
revenue Sales, pricing, churn, ARR, pipeline
operations Deployment, monitoring, incidents, SLAs
security Auth, compliance, audits, vulnerabilities
people Team members, roles, org structure

Cross-domain edges (e.g., engineering → revenue: "enables") are the most valuable — they power the compressed index's cross-reference section that makes Prime beat pure vector search on cross-domain queries.