Built a memory system for chatbots that mimics how humans remember and forget

Sharing a quick breakdown of something I recently built that dramatically improved long-session performance in LLM-based agents.

The challenge => Most chatbots forget everything between sessions — or worse, overload memory with raw logs that aren’t meaningful.

My solution => a structured memory system that’s adaptive, contextual, and forgets what it should. Here’s the architecture:

1. Memory Points - Each user message is turned into a memory point with:

A content summary
A dynamic importance score
A category (like “personal info” or “casual”)
A timestamp
Linked to a user ID

2. Memory Decay - Memories fade over time — but not equally.

Critical info (like names or settings) decays slowly
Small talk fades fast

I use exponential decay with category-based rates.

3. Long-Term Summarization - When a memory fades below a threshold, the LLM generates a long-term summary (e.g., "User likes dark mode"). These are stored as part of the user’s evolving profile.

4. Contextual Retrieval & Conflict Handling - At inference time, the bot pulls in recent memory points + long-term memory. If there are conflicts, it resolves them based on score, recency, and category.

Why it matters:It creates conversations that feel personalized and consistent, without storing everything. It’s lean, adaptive, and avoids token overload.

If anyone else here is building AI agents or tools around personalization/memory — happy to trade notes or dive deeper!

7 comments