# Long term memory **URL:** \ **Category:** LangGraph\ **Tags:** python-help\ **Created:** [August 29, 2025, 7:00pm UTC](https://forum.langchain.com/t/long-term-memory/1408 "2025-08-29T19:00:55Z")\ **Posts on this page:** 3\ **Page:** 1
**Author:** ![zein](https://avatars.discourse-cdn.com/v4/letter/z/a6a055/32.png) [@zein](https://forum.langchain.com/u/zein)\ **Post date:** [August 29, 2025, 7:00pm UTC](https://forum.langchain.com/t/long-term-memory/1408/1 "2025-08-29T19:00:55Z")
1. What’s the best cloud database provider for implementing long-term memory? 2. What are the most efficient strategies currently for applying memory? For example, using a combination of techniques like trim-filtering + base storage + semantic search? ---
**Author:** ![Ziechoes](https://yyz1.discourse-cdn.com/flex007/user_avatar/forum.langchain.com/ziechoes/32/767_2.png) [@Ziechoes](https://forum.langchain.com/u/Ziechoes)\ **Post date:** [September 1, 2025, 12:00am UTC](https://forum.langchain.com/t/long-term-memory/1408/2 "2025-09-01T00:00:26Z")
I’ve been thinking about this too. Most answers focus on _where_ to store memory (cloud DBs, vector search, etc.), but less on _what_ should even be stored. I’ve been testing a different angle: instead of saving everything and then trimming/embedding, only log **failure points** — contradictions, reasoning gaps, unknowns. That way you don’t drown the DB in noise, but you still preserve the moments that actually matter for reasoning. Curious if anyone here has tried “failure-only memory” as a complement to semantic search / base storage? It seems like it could keep long-term memory lean and auditable. ---
**Author:** ![zein](https://avatars.discourse-cdn.com/v4/letter/z/a6a055/32.png) [@zein](https://forum.langchain.com/u/zein)\ **Post date:** [September 8, 2025, 7:59pm UTC](https://forum.langchain.com/t/long-term-memory/1408/3 "2025-09-08T19:59:08Z")
Interesting approach. I’ll give it a try.