Files
ra-h-os/docs/3_context-and-memory.md
T
“BeeRad” 733d1c3407 Initial commit: RA-H Open Source Edition
Local-first knowledge management system with BYO API keys.

Features:
- 3-panel UI (Nodes | Focus | Helpers)
- SQLite + sqlite-vec for vector search
- Agent system (Easy/Hard mode orchestrators)
- Content extraction (YouTube, PDF, web)
- Integrate workflow for connection discovery
- Dimension system with auto-assignment

Tech stack:
- Next.js 15 + TypeScript + Tailwind CSS
- Anthropic (Claude) + OpenAI (GPT) via Vercel AI SDK

Setup:
  npm install && npm rebuild better-sqlite3
  scripts/dev/bootstrap-local.sh
  npm run dev

MIT License
2025-12-15 16:14:28 +11:00

5.5 KiB

Context & Memory

Context Builder

The context builder assembles the information each agent sees during conversations. It creates cacheable blocks (for Anthropic) and structured context for all agents.

Context Structure

Every agent receives:

1. Base Context

  • How nodes, edges, and dimensions work
  • Node reference format: [NODE:id:"title"]
  • If auto-context is enabled, BACKGROUND CONTEXT explains that the 10 most-connected nodes will be listed later
  • Pronouns like "this conversation/paper/video" refer to focused node

2. Agent Instructions

  • Agent-specific system prompt (ra-h, ra-h-easy, wise-rah, mini-rah)
  • Role-specific behavior guidelines
  • Execution approach and response style

3. Tool Definitions

  • Role-specific tools (orchestrator/executor/planner)
  • Each tool's description and parameters
  • Usage guidelines

4. Workflow Definitions (orchestrators only)

  • Available workflows (integrate, etc.)
  • Workflow descriptions and triggers
  • Only shown to ra-h and ra-h-easy

5. Background Context (auto-context hubs) (orchestrators only)

  • Optional block controlled by ~/Library/Application Support/RA-H/config/settings.json
  • Lists the 10 nodes with the highest edge counts (ID + title + edge count)
  • Reminds agents to call queryNodes/getNodesById for full content
  • Ordered by edge count, then most recently updated to break ties

6. Focused Nodes

  • Primary focused node (active tab)
  • Additional focused nodes (other open tabs)
  • 25-word content previews
  • Chunk status and embedding availability
  • Link information

Context Caching

Anthropic (Claude):

  • Explicit cache control blocks (cache_control: {type: 'ephemeral'})
  • Caches base context, instructions, tools, workflows, background context
  • Focused nodes NOT cached (changes frequently)

OpenAI (GPT):

  • Implicit caching based on prefix matching
  • Same block structure for consistency
  • No explicit cache control markers needed

Truncation Strategy

  • Background context: IDs + titles only to keep the block lightweight
  • Focused nodes: 25-word previews
  • Full content access: Agents use queryNodes, getNodesById, searchContentEmbeddings for complete content
  • Chunk status indicator: Shows if embeddings are available (avoid re-extraction)

Memory System (Legacy)

The automatic memory pipeline has been removed. Existing memory nodes remain in the database for archival purposes, but no new ones are created and they are excluded from auto-context. The old files (src/services/memory/**) have been deleted along with the ENABLE_CHAT_MEMORY_PIPELINE toggle. Future improvements should store long-term knowledge as normal nodes with explicit dimensions rather than a background pipeline. 6. Update checkpoint in chat_memory_state

Location:

  • Pipeline: /src/services/memory/synthesis/chatMemoryPipeline.ts
  • Trigger: /src/services/chat/middleware.ts:168-178
  • Extraction: /src/services/memory/synthesis/llmSynthesis.ts

Memory Node Structure

{
  type: 'memory',
  title: 'Insight on [subject]',
  description: 'Multi-line list of facts',
  content: 'Second-person statements combined',
  metadata: {
    category: 'identity' | 'interests' | 'models' | 'preferences' | 'relationships',
    subject_type: 'person' | 'project' | 'concept' | 'resource' | 'organization' | 'workflow',
    source_thread: string,
    source_helper: string,
    importance: 'low' | 'medium' | 'high',
    focused_node_titles: string[],
    canonical_key: string | null,
    subject: string
  }
}

Memory Categories

  • identity - Who you are, your roles, background
  • interests - What you're curious about, learning, exploring
  • models - How you think, mental frameworks, approaches
  • preferences - What you like/dislike, priorities, values
  • relationships - Connections to people, projects, concepts

Subject Types

Memory extraction classifies subjects into these types:

  • person - Individual people (yourself or collaborators)
  • project - Projects, products, or initiatives
  • concept - Ideas, theories, mental models, beliefs
  • resource - Tools, articles, books, references
  • organization - Companies, institutions, communities
  • workflow - Processes, systems, methodologies

Importance Levels

Each memory fact is classified by importance:

  • high - Explicitly marked as priority, main project, strong belief ("is crucial", "must")
  • medium - Normal statements and observations (default)
  • low - Tentative, uncertain, or exploratory statements

Auto-Context Toggle

Auto-context replaces manual pinning. When enabled, BACKGROUND CONTEXT includes the 10 nodes with the highest edge counts.

How it works:

  1. Settings UI exposes a toggle inside the Context tab.
  2. The toggle writes to ~/Library/Application Support/RA-H/config/settings.json.
  3. Context builder and workflows read that file on each request (missing file defaults to false).
  4. When enabled, the following query runs once per request:
SELECT n.id,
       n.title,
       COUNT(DISTINCT e.id) AS edge_count,
       n.updated_at
FROM nodes n
LEFT JOIN edges e ON (e.from_node_id = n.id OR e.to_node_id = n.id)
WHERE n.type IS NULL OR n.type != 'memory'
GROUP BY n.id
ORDER BY edge_count DESC, n.updated_at DESC
LIMIT 10;

Notes:

  • Only orchestrators (ra-h / ra-h-easy) see the block.
  • Titles + edge counts are shown; agents must call queryNodes/getNodesById for content.
  • Users who previously pinned nodes are auto-migrated to autoContextEnabled: true the first time the toggle helper runs and sees legacy pins.