mycelium

Bio-inspired semantic filtering engine. Ecosystem simulation for knowledge quality assessment.

Mycelium takes embedding vectors + text and runs them through a biological ecosystem simulation. Nodes compete, cooperate, merge, and die over 60 ticks. What survives is classified:

| Classification | Meaning | |----------------|---------| | pure | Unique knowledge — survived without merging | | merged | Cluster center — absorbed related content | | loner | Isolated — no social interaction, died alone | | redundant | Duplicate — too similar to another, absorbed early | | dead | Outcompeted — lost in social dynamics |

npm install && npm run build

# Filter from any Qdrant collection
SOURCE_QDRANT_URL=http://localhost:6333 \
SOURCE_COLLECTIONS=my_collection \
VIEW_FORMAT=compact \
npx tsx src/loader/main.ts

No dedicated Qdrant instance needed. Filtering runs entirely in-memory.

Source Qdrant → scroll → slot allocator → IsolatedRunner (in-memory)
                                            ├─ inject (species assign, external weight → w)
                                            ├─ 60 ticks (feelings → action → interaction → death)
                                            ├─ consensus (10 runs, majority vote)
                                            └─ harvest → pure / merged / loner / redundant / dead
                                                           ↓
                                                    digest / manifest / compact / JSON

Each node has:

• Species (anchor, sentinel, herald, summarizer, spore) — determines personality
• Feelings (vigor, dread, kinship, hunger) — computed from self state + environment
• Actions (signal, merge, bequeath, survive) — chosen by personality × feelings
• Resonance — memory of past interactions with each species

Merge direction: higher w always survives as cluster center.