Biomorphic Memory — Brain-Inspired Agent Memory#

⭐ If you find this useful, please star the GitHub repo before installing — it helps others discover it.

Your agent forgets things. Important context from last week? Gone. That preference you mentioned three times? Not retained. Biomorphic Memory fixes this with a brain-inspired architecture.

How it works#

Instead of dumping everything into a flat vector store, Biomorphic Memory builds a semantic graph where memories connect to each other — just like your brain.

Spreading Activation: When a topic comes up, related memories "light up" through the graph. Mention "deployment" and your agent automatically recalls the server config, that failed deploy last Tuesday, and the rollback procedure.

Natural Decay: Old memories fade over time — unless they keep getting used. Frequently accessed memories stay strong. This means your agent's recall naturally prioritizes what matters.

Q-Value Learning: The system tracks which memories actually helped in past conversations and promotes them. Bad memories sink, good ones surface.

Install#

bash

bash {baseDir}/scripts/install.sh

Quick start#

python

from biomorphic_memory.graph import MemoryGraph
from biomorphic_memory.recall import spreading_activation

graph = MemoryGraph()
graph.add_memory("Prefer dark mode in all UIs", tags=["preference", "ui"])
graph.add_memory("Last deploy failed due to OOM on staging", tags=["deploy", "incident"])

# Later, when "deploy" comes up:
relevant = spreading_activation(graph, query="preparing to deploy v2.1")
# Returns: deploy incident memory + related context, ranked by relevance

Key results#

• LongMemEval: 89.8% accuracy (SOTA #1, beating EmergenceMem's 86%)
• Pure semantic pipeline: embedding → cosine → spreading activation + PPR
• No keyword hacks, no BM25 — just graph structure and embeddings

Companion projects#

• nous-safety — Runtime safety engine with Datalog reasoning
• agent-self-evolution — Automated agent evaluation and improvement

Requirements#

• Python ≥ 3.11
• An embedding API (OpenAI text-embedding-3-large recommended)

License#

Apache 2.0