Memory

milo commented

2025-06-07 19:55:09 -04:00

Owner

🧠 Memory Feature Scope – ChromaDB Integration

Goal

Enable long-term memory so the bot can remember facts about users, past conversations, or important events, making interactions feel more alive and personal.

Phase 1: ChromaDB Setup

Add chromadb to requirements.txt
Create a module: memory/chroma_memory.py
Initialize a persistent ChromaDB instance (e.g. ./memory/chroma_db/)
Define default collection: memories
Ensure each memory has:
- id (e.g., user123_2024-06-01)
- user_id or username
- content (the memory fact or message)
- metadata (timestamp, type, source)

Phase 2: Writing Memories

Hook into message handling pipeline (on_message)
Extract key facts/messages (e.g., "I love Apex Legends" → memory)
Store into ChromaDB with metadata
Optional: Use keyword extraction or simple LLM call to summarize what’s worth remembering

Phase 3: Retrieving Memories

Add memory injection to get_ai_response() context:
- On message, retrieve relevant memories via vector similarity (e.g., cosine sim)
- Format as natural-language memory recall (e.g., "Last time you mentioned you love Apex.")
Inject top 3–5 similar memory chunks into the context block

Phase 4: User Control (Optional but recommended)

!remember → Manually add a memory
!memories → List a user’s stored memories
!forget → Wipe memory of a user or phrase

Phase 5: Optimization and Future Steps

Add memory type filtering (e.g., mood, interest, insult, inside joke)
Build sentiment/mood tagging into stored memories
Periodic memory cleanup or summarization
Integrate memory into personality module for adaptive behavior
Use embeddings from LLM instead of default Chroma sentence transformer

Tech Notes

ChromaDB can be run embedded/local, no server needed
Initial vector model: all-MiniLM-L6-v2 or intfloat/e5-small-v2
Consider storing embeddings separately to allow upgrades later

## 🧠 Memory Feature Scope – ChromaDB Integration ### Goal Enable long-term memory so the bot can **remember facts about users**, **past conversations**, or **important events**, making interactions feel more alive and personal. --- ### Phase 1: ChromaDB Setup - [ ] Add `chromadb` to `requirements.txt` - [ ] Create a module: `memory/chroma_memory.py` - [ ] Initialize a persistent ChromaDB instance (e.g. `./memory/chroma_db/`) - [ ] Define default collection: `memories` - [ ] Ensure each memory has: - `id` (e.g., `user123_2024-06-01`) - `user_id` or `username` - `content` (the memory fact or message) - `metadata` (timestamp, type, source) --- ### Phase 2: Writing Memories - [ ] Hook into message handling pipeline (`on_message`) - [ ] Extract key facts/messages (e.g., `"I love Apex Legends"` → memory) - [ ] Store into ChromaDB with metadata - [ ] Optional: Use keyword extraction or simple LLM call to summarize what’s worth remembering --- ### Phase 3: Retrieving Memories - [ ] Add memory injection to `get_ai_response()` context: - On message, retrieve relevant memories via vector similarity (e.g., cosine sim) - Format as natural-language memory recall (e.g., `"Last time you mentioned you love Apex."`) - [ ] Inject top 3–5 similar memory chunks into the context block --- ### Phase 4: User Control (Optional but recommended) - [ ] `!remember` → Manually add a memory - [ ] `!memories` → List a user’s stored memories - [ ] `!forget` → Wipe memory of a user or phrase --- ### Phase 5: Optimization and Future Steps - [ ] Add memory type filtering (e.g., mood, interest, insult, inside joke) - [ ] Build sentiment/mood tagging into stored memories - [ ] Periodic memory cleanup or summarization - [ ] Integrate memory into personality module for adaptive behavior - [ ] Use embeddings from LLM instead of default Chroma sentence transformer --- ### Tech Notes - ChromaDB can be run **embedded/local**, no server needed - Initial vector model: `all-MiniLM-L6-v2` or `intfloat/e5-small-v2` - Consider storing embeddings separately to allow upgrades later