loganionian/0gmem
Long-term conversational memory system for AI agents — cell-based architecture with hybrid BM25 + semantic retrieval. 96% accuracy on LoCoMo benchmark.
What's novel
Long-term conversational memory system for AI agents — cell-based architecture with hybrid BM25 + semantic retrieval. 96% accuracy on LoCoMo benchmark.
Code Analysis
8 files read · 3 roundsA structured long-term conversational memory system for LLMs that encodes entities, temporal relationships, negations, and causal links at ingestion time into a unified multi-graph, then uses multi-strategy retrieval (vector, BM25, entity graph, temporal, hierarchical) with LLM-driven planning, atte
Strengths
Genuinely deep implementation with a coherent write-path (regex NER, temporal parsing, negation detection, causal hints) and read-path (8+ retrieval strategies, attention filter, agentic loop) that are well-integrated. The negation-as-first-class-citizen design threaded through extraction, storage, and retrieval is a thoughtful approach to adversarial robustness that most memory systems lack.
Weaknesses
The core NER and relation extraction is entirely regex-based, which is brittle on unusual phrasing and limits generalization; the LLM-dependent features (query planner, LLM reranker, LLM fact extraction) are all optional and off by default, meaning the 'intelligent' parts require external API calls. The benchmark claims (88.67% on LoCoMo) are not independently verifiable from the code alone and the evaluation harness was not fully readable.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
29
Contributors
1
Files
93
Active weeks
2
Repository
Language
Python
Stars
3
Forks
1
License
MIT