Memory

Employee Profiles

Employee profiles are the persistent memory layer. They give the assistant user-specific context across conversations.

Model

The EmployeeProfile model stores extracted information about each user:

Field	Type	Description
summary	String	LLM-extracted bio / role description
currentProjects	String[]	Active projects the employee works on
preferences	JSON	Communication preferences
frequentIntents	String[]	Commonly used intents
accessLevel	String	Authorization tier (e.g., "standard", "elevated")
metadata	JSON	Custom key-value data
profileVersion	Int	Incremented on each extraction
lastExtractedAt	DateTime	Timestamp of last extraction

One profile per user (one-to-one relationship with User).

How Profiles Are Used

The profile is injected into the LLM context during RAG generation:

• Summary — grounds the assistant in the employee's role.
• Current projects — enables project-aware answers.
• Preferences — shapes response style.

The profile is added to the system prompt before the user's query, which is what gives the assistant memory across conversations.

API

List Profiles

GET /api/v1/employee-profiles

Auth: users see profiles inside their effective department scope; global users may filter by departmentId.

Get Profile

GET /api/v1/employees/:userId/profile

Auth: own profile or a profile inside the caller's effective department scope.

Update Profile

PATCH /api/v1/employees/:userId/profile

Body: any profile fields. Auth: users may update their own profile; canManageEmployeeProfiles is required to update other profiles inside the caller's effective department scope.

Automatic Extraction

The Memory Extraction Worker updates profiles after every N messages (set by MEMORY_EXTRACT_EVERY_N_MESSAGES).

Manual Updates

Users with canManageEmployeeProfiles can edit profiles via the API — useful for seeding initial context or correcting extracted data.

---

Extraction Worker

The extraction worker analyzes conversations and updates employee profiles with extracted facts.

Implementation

src/knowledge/memory-extractor.ts — processMemoryExtractionJob()

Trigger

Every N messages in a conversation (MEMORY_EXTRACT_EVERY_N_MESSAGES, default 10), a BullMQ job is enqueued to memory-extraction.

Pipeline

1. Load profile — fetch the existing EmployeeProfile, create if missing.
2. Fetch messages — pull messages since lastExtractedAt.
3. Format conversation — build a "user: ...\nassistant: ..." transcript.
4. LLM extraction — call gpt-4o-mini (or MEMORY_EXTRACTION_MODEL) with:

   Extract JSON with fields: summary, currentProjects, preferences, frequentIntents, metadata.
   Only include fields with new information.
5. Parse response — extract JSON from the LLM output.
6. Merge with the existing profile:
   • arrays are de-duplicated (union);
   • objects are shallow-merged;
   • strings are replaced.
7. Save — update the profile, increment profileVersion, set lastExtractedAt.

Queue Configuration

Setting	Value
Queue name	memory-extraction
Concurrency	1
Max retries	3
Backoff	Exponential
Worker shutdown	Graceful (SIGTERM)

Configuration

Variable	Default	Description
MEMORY_EXTRACT_EVERY_N_MESSAGES	10	Messages between extractions
MEMORY_EXTRACTION_MODEL	gpt-4o-mini	LLM model for extraction