Conversation Intelligence System

File: assistant/conversation_intelligence.go (~1684 lines)

The Conversation Intelligence system is a 3-phase memory pipeline that transforms raw chat sessions into structured, retrievable knowledge. It powers long-term memory, persona continuity, and autonomous playbook discovery.

Architecture Overview

Chat Session Ends
    ↓
Phase 1: generateConversationSummary()
    → Structured JSONL output (WIS/WIH/WID model)
    → Significance scoring (1-5)
    ↓
Phase 2: RunMemoryDistillation()
    → Processes unprocessed JSONL conversations
    → Routes insights to 3 memory tiers
    → User-isolated processing
    ↓
Phase 3: triggerPlaybookDiscovery()
    → Identifies repeatable workflow patterns
    → Generates playbook candidates
    → Saves to assistant/memory/playbooks/candidates/

Phase 1: End-of-Session Summaries

When a chat session completes, generateConversationSummary() runs asynchronously via goroutine.

Summary Model: openrouter/qwen/qwen3.5-flash-02-23

Output Structure (ConversationSummary):

Field	Description
`Intent`	Primary intent of the conversation
`TechnicalDepth`	Depth level (casual, moderate, expert)
`UserMood`	Detected emotional state
`Significance`	1-5 scale (1=Transient, 5=Foundational)
`RelationshipNotes`	Interpersonal dynamics observed
`ActiveGoals`	Extracted goals (gated to `nikki`, `lia`, `marvin` only)
`WIH/WIS/WID`	What I Heard / What I Said / What I Did model
`ToolsUsed`	Array of `{Name, Purpose}` tool usage
`Outcome`	Resolution status (resolved, partial, unresolved)
`FactsLearned`	New factual knowledge
`PreferencesNoted`	User preference observations
`ErrorsEncountered`	Technical failures
`FollowUp`	Suggested next actions
`Distilled`	Whether this has been through distillation

Retry Logic: 3 retries with exponential backoff (5s, 10s, 20s) on LLM failures.

Storage: JSONL per agent at assistant/memory/{agent}_conversations.jsonl

Excluded Agents (memorySkipAgents): mrtool, analyst, sysadmin, ceo, director, team_leader, employee, elon, projectmanager, queen, bee, worker, scout, soldier, debugger, janitor, memorymaster, memorymanager, accountant, developer, marketing, sales_rep

Phase 2: Memory Distillation

RunMemoryDistillation(targetUser) processes unprocessed conversation summaries and routes insights to three memory tiers:

Memory Tiers:

Tier	File	Max Size	Content
Shared	`assistant/memory/shared_long_term.md`	4,000 chars	Cross-agent knowledge
Agent	`assistant/memory/{agent}_long_term.md`	8,000 chars	Agent-specific expertise
User	`assistant/memory/{user}/{agent}_memory.md`	4,000 chars	Per-user personalization

RBAC Gating: Guest users only see their own user-level memory. Shared and agent memories are restricted to Friend+ roles.

Significance Priority: When truncation is needed, foundational entries (significance 4-5) are preserved first, then utility (3), then recent (1-2).

Concurrency: Each memory file path gets its own mutex via getMemoryFileMutex() to prevent write races.

Distillation Limits: Max 50 conversations per pass (maxJSONLEntriesPerDistillation).

Phase 3: Playbook Discovery

triggerPlaybookDiscovery() analyzes conversation patterns to identify repeatable workflows:

Requirements for a candidate:

4+ distinct tool uses in the pattern
3+ occurrences of the pattern
high or medium confidence only

Output: Candidates saved to assistant/memory/playbooks/candidates/ and notified to Dan via WhatsApp.

Memory Maintenance

ProcessMemoryMaintenance() orchestrates the full cycle:

RunMemoryDistillation() — extract insights from conversations
RunMemoryReflection() — "The Big Tidy": retroactively assigns significance, merges redundant facts, prunes low-signal entries, converts to WIS/WIH/WID format

Config:

Constant	Value	Purpose
`ConversationSummaryModel`	`openrouter/qwen/qwen3.5-flash-02-23`	LLM for summarization
`playbookLookbackDays`	14	How far back playbook discovery looks
`maxJSONLEntriesPerDistillation`	50	Max conversations per distillation pass
`maxSharedMemoryChars`	4,000	Shared memory budget
`maxAgentMemoryChars`	8,000	Agent memory budget
`maxUserMemoryChars`	4,000	User memory budget

Key Functions

Function	Purpose
`generateConversationSummary()`	Phase 1: LLM-powered structured summary of chat session
`RunMemoryDistillation()`	Phase 2: Process conversations, route to memory tiers
`RunMemoryReflection()`	Phase 2+: Retroactive significance, dedup, WIS/WIH/WID conversion
`BuildMemoryContext()`	Assemble 3-tier memory context for agent prompts
`triggerPlaybookDiscovery()`	Phase 3: Find repeatable workflow patterns
`shouldSkipMemory()`	Check if agent is excluded from memory saving
`loadMemoryFile()`	Load + truncate memory with significance prioritization
`lockedAppendToFile()`	Thread-safe file append with per-path mutex