A drop-in memory layer for AI chatbots. ChatSorter scores every message, extracts structured facts, and returns only the context that's relevant right now — not everything ever stored.
# 1. Store every message requests.post("api.chatsorter.com/process", json={ "chat_id": user_id, "message": user_message }) # 2. Retrieve relevant context r = requests.post("api.chatsorter.com/search", json={ "chat_id": user_id, "query": user_message }) # What comes back # [{ "summary": "User is allergic to peanuts", # "source": "structured", "score": 0.94 }]
Most memory tools store everything and dump it all into context. ChatSorter scores, filters, and returns only what's relevant right now.
Every message scored 1–10. "I love pizza" scores 4. "I'm allergic to peanuts" scores 10. Only high-signal messages get extracted permanently.
Structured key/value facts with confidence scores. "My dog is named Max" becomes { pet: "Max", confidence: 0.95 }. Not a paragraph of text.
Every N messages compressed into a third-person summary. Your model sees 20 summaries instead of 400 raw messages. Token costs drop fast.
Query returns only what matters for the current message. Semantic similarity + importance weighting + recency decay — all combined into one score.
When someone says "I moved to SF," the system updates location — not appends it. Volatile facts like health conditions are tracked separately with staleness detection.
Two API calls. POST /process to store a message. POST /search to retrieve relevant context. Works with any Python backend, any LLM, any stack.
Most tools have one layer — a database. ChatSorter has three, each doing a specific job.
Last N messages in a rolling window. No LLM involved. Zero latency. Keeps the model aware of recent context without any processing overhead.
Every batch of N messages compressed into a single third-person summary via local LLM inference. Stored with importance scores and timestamps. Decays over time.
High-signal messages parsed into typed key/value facts — name, job, allergies, pets, preferences. Confidence-scored. Volatile facts (health, lawsuits) tracked separately with staleness detection.
ChatSorter is built specifically for chatbot memory — not general-purpose vector storage.
| Feature | Mem0 | Supermemory | ChatSorter |
|---|---|---|---|
| Confidence scores on facts | ✗ | ✗ | ✓ |
| Importance-gated extraction | ✗ | ✗ | ✓ |
| Volatile fact tracking | ✗ | ✗ | ✓ |
| Bring your own vector DB | ✗ | ✗ | ✓ |
| Local inference option | ✗ | ✗ | ✓ |
| Free tier | Limited | 1M tokens/mo | Full beta access |
| Target use case | Enterprise | Full platform | Chatbot memory engine |
Free beta access. No credit card. Two endpoints away from working memory for your chatbot.
Start building →