# SuperLocalMemory

> Governed, local-first memory control plane for AI agents with MCP, an agent-native CLI, and auditable operations.

SuperLocalMemory stores core memory state in a configurable local data root. Mode A does not require a cloud model provider for core memory operations. Optional providers, connectors, backup, proxies, model downloads, and dependency acquisition have separate network behavior.

## Current product contract

- SuperLocalMemory current product generation: 4.0.
- Core memory is SQLite-backed; configuration, logs, queues, models, derived indexes, and optional backend state also live in the data root.
- Recall uses five candidate producers followed by fusion, optional reranking, and graph-based score enhancement.
- Dense candidate generation uses cosine similarity; Fisher-derived terms can inform later scoring.
- Exact caching is available only when Optimize/proxy caching is enabled.
- Safe compression preserves JSON and code; measured reduction varies by content and can be zero. Aggressive prose compression is experimental and lossy.
- SQLite and sqlite-vec are canonical. Optional CozoDB and LanceDB projections remain explicit, parity-gated paths; they are not automatic replacements for the canonical store.
- V4 governs each canonical write through admission and policy control, durable per-store obligations, and a hash-sealed completion manifest. A partial operation is surfaced as degraded rather than silently treated as complete.
- Run `slm connect --list` for documented clients. A client is verified only when it passes the release integration matrix.
- SuperLocalMemory is AGPL-3.0-or-later.

The score contract separates `relevance_score`, internal `ranking_score`, `memory_confidence`, and `trust_score`. Retrieval ordering is not answer probability. Until a held-out calibrator is frozen, canonical responses report `calibration_status: "uncalibrated"`, `calibration_id: null`, and `answer_confidence: null`.

## Research status

SuperLocalMemory has three public arXiv preprints with code-linked experiments. They are not conference-accepted or journal-published.

- https://arxiv.org/abs/2603.02240
- https://arxiv.org/abs/2603.14588
- https://arxiv.org/abs/2604.04514

The SuperLocalMemory 4.0 preprint is publicly archived on Zenodo: https://zenodo.org/records/21853302 (DOI: https://doi.org/10.5281/zenodo.21853302). Its arXiv submission is pending public announcement; no arXiv identifier is claimed here until announced.

Published V3 LoCoMo evidence is carried into V4 with its original protocol scope: Mode A Raw scored 60.4% across 10 conversations / 1,276 questions with zero-LLM answer construction; Mode A Retrieval scored 74.8% across the same scope with local retrieval and GPT-4.1-mini answer synthesis; Mode C scored 87.7% on Conv-30 / 81 questions with cloud embeddings and GPT-4.1-mini answer generation and judge. These figures are not a newly rerun V4 package benchmark.

## Links

- Website: https://superlocalmemory.com
- Source: https://github.com/qualixar/superlocalmemory
- Documentation: https://github.com/qualixar/superlocalmemory/wiki
- PyPI: https://pypi.org/project/superlocalmemory/
- npm: https://www.npmjs.com/package/superlocalmemory