An end-to-end agentic memory framework built around CogniGraph, a lightweight hierarchical graph that uses entities and relations as semantic indexing layers. Targets efficient long-term reasoning: keep the memory graph small and the retrieval cheap while preserving multi-session recall. Published as an arXiv 2025 paper with public code.
- Storage
- CogniGraph — a hierarchical graph of entities and relations serving as semantic indexing layers, backed by dense embeddings (BGE-M3 in the paper's evaluation).
- Retrieval
- Temporal- and hierarchy-aware search over the CogniGraph with integrated reranking, aiming for adaptive, coherent knowledge retrieval at low token cost.
- Self-host
- Self-host: moderate
- License
- unlicensed
- Pricing
- Research code on GitHub; no LICENSE file (all rights reserved by default). No hosted or paid tier. · Free / OSS
- GitHub stars
- 47
- Last release
- —
- Last commit
- 2026-01-06
- First catalogued
- 2026-06-28
Strengths
- Graph-structured memory (CogniGraph) with temporal + hierarchy-aware search and reranking
- Paper reports strong LoCoMo and LongMemEval accuracy at low token cost across two backbones
- Its evaluation provides a clean head-to-head re-run of Mem0/A-Mem/Zep (useful as an independent baseline source)
Watch out
- No LICENSE file in the repo — code is public but all-rights-reserved by default; not legally reusable until licensed
- Research-grade maturity: ~47 GitHub stars, no tagged releases, last push 2026-01-06
- Headline scores are self-reported (authors' own system); independent reproduction pending
Best for
- Researchers comparing graph-based agentic memory under a controlled, efficiency-focused evaluation
Benchmark results
| Benchmark | Value | Backbone | Trust | Source |
|---|---|---|---|---|
| longmemeval | 73.8 accuracy | gpt-4o-mini | Self-reported | LiCoMemory (Huang et al., HKUST/Huawei/CUHK-SZ/WeBank) ↗ |
| locomo | 67.2 accuracy | gpt-4o-mini | Self-reported | LiCoMemory (Huang et al., HKUST/Huawei/CUHK-SZ/WeBank) ↗ |
| longmemeval | 69.2 accuracy | Llama-3.1-70B-Instruct-Turbo | Self-reported | LiCoMemory (Huang et al., HKUST/Huawei/CUHK-SZ/WeBank) ↗ |
| locomo | 62.99 accuracy | Llama-3.1-70B-Instruct-Turbo | Self-reported | LiCoMemory (Huang et al., HKUST/Huawei/CUHK-SZ/WeBank) ↗ |
Sources
- LiCoMemory: Lightweight and Cognitive Agentic Memory for Efficient Long-Term Reasoning (paper)
- LiCoMemory repo (README, code) (vendor)
- GitHub API metadata (47 stars, no license file, pushed 2026-01-06) (third-party)
Last verified 2026-06-28 · updated by discover-frameworks