Features
Everything an agent needs to remember — nothing it needs to leak.
The Community Edition below is the complete framework. Pro adds a commercial license and enterprise controls on top.
Memory-palace structure
Rooms → loci → chunks. Agent memory gets architecture: workspaces map to rooms, topics to loci, and every memory is a placed, retrievable chunk with provenance.
Hybrid retrieval, fused
BM25 full-text (SQLite FTS5), dense vector search, and graph traversal run per query and merge with reciprocal rank fusion — lexical precision plus semantic recall.
Three-layer forgetting
L1 time-based decay, L2 similarity interference, L3 budget eviction. Memories degrade gracefully — FP32 → FP16 → INT8 → BINARY → released — under an explicit storage budget.
On-device embeddings
Embedding inference runs locally via ONNX Runtime. No API keys, no per-query cost, no third-party sees your context. Device profiles tune it from laptops to servers.
Single-file SQLite store
The entire palace — content, vectors, graph, feedback log — lives in one SQLite file. Back it up with cp. Inspect it with any SQLite client.
Token-budgeted assembly
One call retrieves, ranks, and assembles context to a token budget — ready to paste into a prompt. The Twin-Force Controller adapts retrieval behaviour from interaction signals.
Conversation memory
Store full conversation turns with a single API call; retrieved-chunk feedback is logged automatically so useful memories strengthen and noise decays.
Graduated deletion (RTBF-capable)
Deletion is a first-class lifecycle stage, not a missing feature. Pro adds the compliance toolkit: atomic erasure with a tamper-evident, hash-chained audit trail.
REST API + live dashboard PRO
A FastAPI server exposes search, store, assemble, turn, recall, and feedback endpoints — with an effectiveness dashboard that works fully offline (Chart.js is vendored, CSP-strict).
See it running
The effectiveness dashboard.
Palace topology, Twin-Force Controller state, retrieval benchmarks, memory budget, and live search — served by your own machine at localhost:8848/dashboard (Lummenna Pro).
Comparison
Against common vector stores.
| Feature | Lummenna | Chroma | FAISS | pgvector |
|---|---|---|---|---|
| BM25 lexical search | Built-in (FTS5) | No | No | No |
| Hybrid retrieval (lexical + vector) | RRF fusion | Manual | No | Manual |
| Memory lifecycle (decay + eviction) | 3-layer | No | No | No |
| Runs fully offline | Yes (ONNX) | Partial | Yes | Yes |
| Single-file embedded storage | Yes (SQLite) | No | Index files only | Postgres server |
| LangGraph checkpoint saver | Built-in | No | No | No |
| MCP server | Built-in | No | No | No |
| Graduated deletion (RTBF-capable) | Yes | DELETE only | Delete only | DELETE only |
| Token-budgeted context assembly | Yes | No | No | No |