Context scoping, sensitivity, and entity linking for knowledge

[kovtcharov]

Summary

Implement the three cross-cutting features that make memory useful for a PC assistant: context scoping (work vs personal), sensitivity classification (private data protection), and entity linking (people, apps, services).

Spec: docs/spec/agent-memory-architecture.md → "Context Scoping", "Sensitivity Classification", "Entity Linking"

Depends on: #542 (MemoryStore)

Context Scoping

Separate knowledge by life domain. System prompt only loads active context + global.

• context column on knowledge + conversations (default: 'global')
• init_memory(context="work") sets active context
• set_memory_context("personal") switches mid-session
• System prompt: global + active context items
• Dedup scoped to context
• remember() defaults to active context (overridable)
• recall() searches all contexts by default (filterable)

Predefined: global, work, personal. User-defined: any string (project-x, learning, etc.)

Sensitivity Classification

Protect private data from leaking into system prompt or logs.

• sensitive INTEGER column (0/1, default 0)
• sensitive=1: excluded from system prompt, returned only via explicit recall()
• Tool history: args redacted to keys only for sensitive operations
• Dashboard: 🔒 badge, content blurred until clicked

Entity Linking

Associate knowledge with people, apps, services using lightweight type:name tags.

• entity TEXT column (e.g., 'person:sarah_chen', 'app:vscode', 'project:gaia')
• get_by_entity() — indexed lookup for "everything about X"
• recall(entity="person:sarah_chen") — all facts about Sarah
• Multiple knowledge entries can share an entity (builds profile over time)
• No separate entity table — denormalized tag on knowledge rows

Acceptance Criteria

• Context column on knowledge and conversations
• System prompt filters by active context + global
• Sensitive items excluded from system prompt
• Sensitive items returned via explicit recall
• Entity linking works with type:name convention
• get_by_entity returns all linked knowledge
• Dedup respects context boundaries
• Dashboard filters work for all three features