An external hippocampus for your chatbot

What it does MemoryBear is a FastAPI service that ingests conversations and documents, extracts structured triples into a Neo4j knowledge graph, and surfaces them through a hybrid Elasticsearch-and-vector search layer. It wraps this in a biologically-themed lifecycle where memories are reinforced through use and pruned when they decay.

The interesting bit The project treats forgetting as a first-class feature rather than a bug. A scheduled reflection engine runs daily consistency checks, and a synaptic-pruning model slowly drives unused knowledge through dormancy into clearance. That is the central conceit separating it from simpler vector stores.

Key highlights

• Graph storage in Neo4j covering 12 relationship types (hierarchical, causal, temporal, logical, etc.)
• Hybrid keyword + semantic search claiming 92% accuracy and a 35% improvement over single-mode retrieval
• Explicit memory lifecycle with strength decay and a three-stage dormancy → decay → clearance pipeline
• Celery-based async architecture splitting memory, document, and periodic tasks across separate queues
• FastAPI layer with dual endpoints: JWT-backed management and API-key service access

Caveats

• Running it requires orchestrating PostgreSQL, Neo4j, Redis, and Elasticsearch in addition to the provided Docker Compose file, which only covers the API and workers.
• Benchmark charts claim superiority over Mem0, Zep, and LangMem, but the README offers little detail on the underlying tasks or evaluation protocol beyond the metrics shown.

Verdict Worth a look if you are building long-running agents or customer-facing chat systems that need persistent, evolving memory with explicit pruning. If you wanted a lightweight in-process memory buffer, this is overkill.