widemem: open-source memory layer for LLMs with importance scoring, decay, and conflict resolution

Show and Tell
1

Hi everyone, sharing an open-source project I’ve been working on.

widemem is a Python memory layer for LLMs that goes beyond basic vector search. Most memory systems treat all facts equally and let contradictions accumulate silently. widemem tries to fix that.

What it does

  • Importance scoring: each extracted fact gets a 1-10 score. Retrieval ranks by a weighted mix of similarity, importance, and recency
  • Time decay: configurable exponential/linear/step decay. Old trivia fades, critical facts stick
  • Batch conflict resolution: “I moved to Paris” after “I live in Berlin” gets resolved in a single LLM call instead of storing both
  • YMYL prioritization: health, legal, and financial facts get higher importance floors and decay immunity
  • Hierarchical memory: facts roll up into summaries and themes with automatic query routing
  • MCP server: works as a Model Context Protocol server for Claude Desktop and other MCP clients

Runs fully local

Works with Ollama + sentence-transformers + SQLite + FAISS out of the box. No cloud, no API keys needed. Also supports OpenAI, Anthropic, and Qdrant if you want.

pip install widemem-ai

140 tests passing. Apache 2.0.

Would love feedback, especially from anyone working on agent memory or RAG pipelines. What approaches are you using for memory in your projects?

1 Like