Next-generation AI memory system inspired by hippocampal memory encoding and neocortical knowledge consolidation. Spans the full knowledge lifecycle: perception → extraction → association → forgetting. LLM-driven extraction converts conversations into structured entity-relationship triples stored in a Neo4j knowledge graph, while a parallel vector store enables hybrid semantic+keyword retrieval. A biologically-motivated forgetting engine (dormancy → decay → clearance) prunes low-value knowledge automatically.
- Storage
- Dual-store architecture: Neo4j knowledge graph stores extracted entity-relationship triples with temporal anchors and provenance; a vector store (BERT embeddings) holds parallel semantic representations for similarity search. LLM extraction at write time converts raw text into structured triples and syncs them to both stores. Each knowledge item carries a dynamic memory-strength score updated by usage frequency and association activity.
- Retrieval
- Dual-engine hybrid retrieval: BERT-based semantic vector search (synonym and implicit-intent aware) expands the candidate space; BM25 keyword search then performs precise filtering over candidates. The two signals are fused via RRF. Interactive knowledge-graph visualisation available for human inspection. Self-reported internal benchmark: 92% retrieval accuracy vs 57% single-mode baseline (vendor measurement, untracked benchmark — config not published).
- Self-host
- Self-host: heavy
- License
- Apache-2.0
- Pricing
- Open-source Apache-2.0. Self-hosting requires Neo4j (non-trivial to operate). A commercial cloud service appears to exist at memorybear.ai; pricing was not captured. · Free + paid
- GitHub stars
- 4,661
- Last release
- 2026-06-26
- Last commit
- 2026-06-27
- First catalogued
- 2026-06-28
Strengths
- Full knowledge lifecycle with biologically-motivated forgetting: dormancy → decay → clearance keeps memory collections pruned rather than accumulating stale data
- Dual Neo4j + vector store architecture supports rich relational queries (graph traversal) alongside semantic similarity — stronger than pure vector-only memory
- Multimodal affective intelligence path (arXiv:2603.22306) extends to audio, image, and emotion-aware memory for assistants
- 4661 stars, actively maintained with frequent releases (v0.3.9, 2026-06-26); Apache-2.0
Watch out
- Neo4j dependency makes self-hosting heavier than most catalog entries — production deployments need a managed or self-hosted Neo4j instance
- Organisation is a Chinese entity (Suanmo Suanyang Technology / RedBear AI); evaluate long-term maintainership and supply-chain considerations
- Self-reported 92% retrieval accuracy is on an untracked internal benchmark with no published config — treat as directional, not comparable to LoCoMo/LongMemEval results
- Cloud commercial tier (memorybear.ai) pricing not publicly listed
Best for
- AI assistants and agents that need rich relational knowledge management — especially where entity relationships, temporal tracing, and automated forgetting matter more than pure vector recall speed
Benchmark results
No sourced results yet.
Sources
- MemoryBear README (vendor)
- Memory Bear AI Memory Science Engine for Multimodal Affective Intelligence (arXiv) (paper)
- GitHub API repo metadata (4661 stars, Apache-2.0, v0.3.9 release) (third-party)
Last verified 2026-06-28 · updated by discover-frameworks