Lummenna

Estrutura de Memória e Contexto que organiza memórias em uma estrutura de "palácio da memória" (salas, loci, chunks), usa recuperação híbrida, gerencia o decaimento de memória e é executada inteiramente no hardware do cliente, sem dependência de nuvem. Os embeddings são executados localmente por meio do ONNX Runtime, o armazenamento é SQLite em arquivo único e as integrações incluem um checkpoint saver do LangGraph, um adaptador LangChain, um servidor MCP e uma API REST.

Ver em lummenna.com

Documentação

Crystalline memory field / local, persistent, illuminated

E2E memory / Recall@1 0.964

28 queries across 7 synthetic sessions; MiniLM embedder.

BEIR subset / nDCG@10 0.439

Hybrid macro mean across 5 datasets, each sampled to 500 passages and 20 queries.

Leaderboard status Unranked

No full-corpus BEIR/MTEB submission or external memory-system comparison is committed.

01 / Structure

Rooms, loci, chunks

Memories carry location, provenance, validity, value, and retrieval state instead of living in an undifferentiated vector collection.

Community02 / Retrieval

Signals, then fusion

SQLite FTS5, dense vectors, optional graph traversal, and reciprocal rank fusion form the retrieval path.

Community03 / Context

Budgeted assembly

Retrieved memories are ranked and assembled into bounded prompt context with explicit truncation behavior.

Community04 / Lifecycle

Retention has pressure

Time decay, interference, and budget-driven precision changes prevent an unbounded store from masquerading as useful memory.

Community05 / Interfaces

Use the same palace

Python, MCP, FastAPI, LangChain, and LangGraph adapters connect agents to the same underlying memory model.

Community06 / Controls

Enterprise boundaries

Encryption at rest, tenant isolation, compliance operations, and advanced P2P controls are license-enforced Pro capabilities.

Pro01 / Organize

Place memory

Store a chunk in a named room and optional locus, with provenance and temporal validity recorded beside it.

02 / Retrieve

Fuse evidence

Lexical and semantic candidates are combined, reranked, tenant-scoped, and reinforced when used.

03 / Maintain

Release deliberately

Scheduled decay and budget passes alter value and precision according to explicit, inspectable rules.

Available integrations

Run the observatory beside your agent.

The direct Community distribution includes a package, corresponding source, and checksum-pinned sidecar recipe without source-control access.