Long-term memory for LLMs is a solved problem in the sense that several libraries exist. What isn't solved, according to GitHub user remete618, is the quality of what gets remembered. Widemem, an open-source Python library released under the Apache 2.0 license, is built around three specific failures the author identified in existing systems: all memories carry equal weight regardless of importance, contradictory facts accumulate without resolution, and nothing protects critical information from time-decay functions that quietly deprioritize it. The library ships as widemem-ai on pip, requires no external services, defaults to SQLite for persistence and FAISS for vector search, and supports Python 3.10 and above with 140 passing tests.
Importance scoring sits at the center of the design. Each memory receives a 1–10 score at extraction time, and retrieval ranks results by combining semantic similarity, importance, and recency into a single weighted score. Memory is organized across three layers — raw facts, summaries, and themes — so a query like "where does Alice live?" pulls from facts while "tell me about Alice" returns a summary, neither requiring an extra API call.
Conflict resolution works in batch: all incoming facts are evaluated against related existing memories in a single LLM call rather than one call per fact. The author argues this cuts cost and avoids the semantically incoherent resolutions that can emerge when individual calls don't account for each other.
The YMYL subsystem — Your Money or Your Life — is the most opinionated piece. Health, legal, and financial facts get decay immunity, higher importance floors, and forced contradiction detection. The practical motivation is explicit: a memory system that forgets a user's medication dosage because time-decay deprioritized it creates real liability. For everything outside YMYL, decay is configurable as exponential, linear, or step functions.
LLM backends include OpenAI, Anthropic Claude, and Ollama; sentence-transformers handles embeddings for air-gapped deployments; Qdrant is available as a vector store for production use.
The project is early — a roadmap item for self-supervised extraction, which would use LLM-based extraction to distill a smaller local model over time, hasn't shipped yet — but it addresses a real gap. Commercial memory products like Mem0 and Zep add memory infrastructure with managed dependencies and opaque internals. Widemem is built for developers who want explicit control over importance weighting, decay curves, and conflict resolution without routing user data through a third-party service. Source code is at github.com/remete618/widemem-ai.