I built a memory layer for Convex apps and just shipped a feature that lets your AI do RLM
If you've ever plugged an LLM into your app, you've hit this wall: you embed your docs, search returns the "top 5 most similar chunks," and the AI answers based on those 5 pieces. Works fine for "what's our refund policy?" but completely falls apart for "compare how we describe pricing across all our docs and flag any inconsistencies."
The top-5 approach can't do that. You need the AI to see everything.
Memcity is a Convex component that handles the full memory/search stack:
- Hybrid search (vector + keyword) with a 16-step pipeline, completely configurable
- Automatic knowledge graphs (entities + relationships)
- Per-user episodic memory (your app remembers each user's preferences, context, history)
- 25+ file types (PDF, images, audio, video -- all processed automatically)
But the thing I'm most excited about is RLM (Recursive Language Models).
Instead of "find top 5 chunks and hope for the best," RLM loads ALL your data into a sandbox as a JavaScript variable. Then an LLM writes code to explore it -- filtering, grouping, calling sub-LLMs to analyze sections, iterating until it has an answer. The data never goes into the LLM's context window. The LLM navigates it programmatically.
This means your app can answer questions that require reading across hundreds of documents, not just 5 chunks.
It's all Convex-native except the RLM which uses sandboxes. No Pinecone. No separate vector DB. Everything runs on Convex tables, vector indexes, and actions. One component,
The community tier is free (Apache 2.0) and includes hybrid search, chunking, and the core pipeline.
Pro ($79 one-time) adds knowledge graphs, file processing, and episodic memory.
Team ($179 one-time) adds RLM and enterprise features.
The search and AI chat at memcity.dev are powered by Memcity.
Docs: https://memcity.dev/docs/getting-started
RLM deep dive: https://memcity.dev/docs/rlm
Would love to hear what you think, especially if you're building anything with AI + Convex.
The top-5 approach can't do that. You need the AI to see everything.
Memcity is a Convex component that handles the full memory/search stack:
- Hybrid search (vector + keyword) with a 16-step pipeline, completely configurable
- Automatic knowledge graphs (entities + relationships)
- Per-user episodic memory (your app remembers each user's preferences, context, history)
- 25+ file types (PDF, images, audio, video -- all processed automatically)
But the thing I'm most excited about is RLM (Recursive Language Models).
Instead of "find top 5 chunks and hope for the best," RLM loads ALL your data into a sandbox as a JavaScript variable. Then an LLM writes code to explore it -- filtering, grouping, calling sub-LLMs to analyze sections, iterating until it has an answer. The data never goes into the LLM's context window. The LLM navigates it programmatically.
This means your app can answer questions that require reading across hundreds of documents, not just 5 chunks.
It's all Convex-native except the RLM which uses sandboxes. No Pinecone. No separate vector DB. Everything runs on Convex tables, vector indexes, and actions. One component,
The community tier is free (Apache 2.0) and includes hybrid search, chunking, and the core pipeline.
Pro ($79 one-time) adds knowledge graphs, file processing, and episodic memory.
Team ($179 one-time) adds RLM and enterprise features.
The search and AI chat at memcity.dev are powered by Memcity.
Docs: https://memcity.dev/docs/getting-started
RLM deep dive: https://memcity.dev/docs/rlm
Would love to hear what you think, especially if you're building anything with AI + Convex.
Join the Convex Discord! Explore Convex, the backend platform that keeps your app in sync.
15,590Members
Resources
Recent Announcements
Similar Threads
Was this page helpful?
