[Feature] Deep Remory Integration: Auto-search and auto-store with graceful fallback

[randomm]

Summary

Replace the current prompt-driven remory approach with deep system integration. Auto-search before tasks, auto-store after completion, with graceful fallback when remory is unavailable.

Background

Research findings (Jan 29, 2026):

• Current manual approach produces HIGH quality memories (47 stored, avg 468 chars)
• Agents store synthesized conclusions, not reasoning noise ✅
• Gap: Manual process means agents may miss relevant context or forget to store

Requirements

1. Remove Existing Stub Implementation

• Remove remory_persist_thinking config flag (unused)
• Clean out any stubbed/unused remory code paths
• Keep the core memory/remory.ts API

2. Pre-Task Auto-Search

When a task/subagent is dispatched:

const memories = await remory.search(taskDescription, { user_id, limit: 5 })
const relevant = memories.filter(m => m.distance < 0.40) // Confidence threshold
// Inject into task system prompt as context

3. Post-Task Auto-Store

When a task completes successfully:

// Generate one synthesized summary (NOT reasoning dump)
const summary = `Task: ${task.description}\nOutcome: ${task.status}\nApproach: ${task.summary}`
await remory.add(summary, {
  user_id,
  metadata: {
    type: inferType(task), // 'root_cause' | 'investigation' | 'fix' | 'pattern'
    issue: task.issueId,
    agent: task.agentType,
    date: new Date().toISOString()
  }
})

4. Graceful Fallback (CRITICAL)

If remory is not available (not installed, socket unavailable, errors):

• Log warning once per session, not per operation
• Continue without memory features — don't block or crash
• All remory calls must be wrapped in try/catch with fallback

async function safeRemorySearch(query: string): Promise<Memory[]> {
  if (!remoryAvailable) return []
  try {
    return await remory.search(query, { user_id, limit: 5 })
  } catch (e) {
    markRemoryUnavailable()
    log.warn("Remory unavailable, continuing without memory features")
    return []
  }
}

5. Structured Metadata

Add consistent metadata tagging:

interface MemoryMetadata {
  type: 'root_cause' | 'investigation' | 'security' | 'fix' | 'pattern' | 'state'
  issue?: string      // GitHub issue number
  agent?: string      // Agent type that created it
  files?: string[]    // Related file paths
  date: string        // ISO timestamp
}

Files to Modify

• packages/opencode/src/memory/remory.ts — Add graceful fallback wrapper
• packages/opencode/src/config/config.ts — Remove remory_persist_thinking stub
• packages/opencode/src/session/ — Add pre-task search injection
• packages/opencode/src/task/ — Add post-task auto-store
• Agent prompt files — Remove manual remory instructions (now automatic)

Acceptance Criteria

• Stubbed/unused remory code removed
• Pre-task auto-search injects relevant memories into context
• Post-task auto-store saves synthesized summary with metadata
• Graceful fallback when remory unavailable (no crash, warning logged once)
• Works transparently — agents don't need to call remory manually
• Existing manual remory search/add tools still available for explicit use

Non-Goals

• ❌ Storing reasoning blocks (confirmed low value, high noise)
• ❌ Storing every tool result
• ❌ Making remory a hard dependency

References

• Research: Remory value analysis (Jan 29, 2026)
• Current memories show high quality when synthesized
• ../remory/ — Remory codebase for reference
• ~/.remory/remory.db — SQLite database

[randomm]

Implemented in 082078e: Deep remory integration with auto-search/store and graceful fallback