Vibeleaderboard
Index / tool
Visit github.com
Category
AI Agents
Rank
No. 1180Tools index

Previous survey · No. 1167 ·

Pricing
Open Source
Type
TOOL
Use case
Data, Retrieval & Knowledge
Interfaces
Agent Skill / Plugin
GitHub
278 stars
Latest release
v4.2.19
Date

About

SuperLocalMemory is an open-source, local-first memory control plane for AI agents and teams, with governed memory and a linked arXiv paper. The repository includes plugins for several agent tools, including Claude, Codex, Copilot, and Hermes.

What it does

A background service on your own computer that stores facts, decisions, screenshots and PDFs, then answers questions about them for any coding assistant that speaks the Model Context Protocol (MCP) or can run a shell command. A query runs through several search methods at once (meaning, keywords, time references, linked facts), merges the lists, reranks them, and then decides whether anything found actually answers it. When nothing does, it says so in a fixed sentence instead of returning the nearest weak match. By default, saving and searching call no hosted model; two optional modes add a local or cloud model.

Why it's ranked here

The recall gate is the differentiator, and the project publishes its limits: on its own held-out set it answered 77 of 77 answerable questions and returned nothing for 152 of 156 nonsense ones, but caught only 7 of 50 plausible questions with no stored answer. That candour is rare among memory servers. Write receipts, undoable corrections, point-in-time filters and hard token caps on recall output back that up. It ships on PyPI and npm, registers as an MCP server, runs on Windows, Linux and Apple Silicon Macs, and declares itself beta. The AGPL license and a very heavy dependency stack keep it from being a default pick.

What's good

Recall protects the context window: it returns a one-line glance, full facts, or facts with surrounding context, under a token cap that is never exceeded, and an agent can drill into one result on demand. Corrections replace a fact without deleting it and can be rolled back. Time filters ask what was true, or what was known, at a given moment. Under load it refuses to hand back a degraded keyword guess and returns a busy code with a retry hint instead. The remote reranker path redacts recognised secrets, requires HTTPS off-host and refuses redirects. A client can be limited to 19 of the 110 tools.

Tradeoffs

The install is heavy. Every dependency is hard-pinned, and the base package pulls in PyTorch, Transformers, LightGBM and a prose-compression library whose model is roughly 560MB on first setup. On Apple Silicon the setup wizard installs a 1.1GB answer-check model by default. Picture and PDF search adds another 1.5GB and needs 16GB of memory. Intel Macs are unsupported. Cloud provider keys sit in plaintext in a permission-restricted config file. The bundled reranker is English-only, and a remote embedding endpoint gets no secret filtering, unlike the remote reranker. AGPL-3.0 matters if you embed it in a service. The default tool set exposes 58 tools.

How to use it well

Best for one developer running several coding assistants on one machine who wants shared, inspectable memory without a hosted service. Install with pip or npm, pick the local mode, and connect one assistant first. Set the MCP tool set to core to keep the client's tool list short. Have the agent start at the glance depth and expand only what it needs. Treat the closest-memories tier as a hint, not an answer, since the gate misses most plausible questions memory cannot answer; turn on the answer check if that matters. It is not a hosted, multi-tenant backend.

Technical notes+

pyproject.toml hard-pins every runtime dependency (torch==2.13.0, transformers==5.10.4, sentence-transformers==5.6.1, llmlingua==0.2.2), with comments warning that transformers 5.17+ removes a function the nomic-embed remote modeling code calls, and pins pycozo to 0.7.6 only on linux aarch64 because the Cozo on-disk format is not forward-compatible; the classifier is Development Status 4 - Beta, coverage fail_under is 0, and default pytest addopts deselect the slow, ollama, benchmark and native markers. src/superlocalmemory/__init__.py unconditionally sets KMP_DUPLICATE_LIB_OK=TRUE and defaults OMP_NUM_THREADS to 2 at import, to stop OpenMP aborts when torch, scikit-learn and LightGBM share a process, and reads package metadata (without importing) to warn if sentence-transformers or onnxruntime drift from their pins. server.json registers stdio packages for uvx and npx, both launched with the mcp argument; docs/mcp-tools.md adds an HTTP daemon transport on 127.0.0.1:8765 and SLM_MCP_PROFILE tool sets (core 19, default full 58, whole 110, unknown values are an error). docs/ARCHITECTURE.md makes SQLite canonical with an FTS projection committed in the same transaction as the fact, moves writes through raw, queryable, enriching and complete phases, and fuses five candidate producers with Reciprocal Rank Fusion before optional reranking. docs/recall.md documents the support gate as a fitted weighted sum, its held-out figures, the SLM_RECALL_NO_SUPPORT_GATE switch and a roughly twelve-second wait before RECALL_BUSY (CLI exit 75). docs/configuration.md states provider keys are stored in plaintext with a 0600 write, and that the OpenAI-compatible embedding path sends raw text with no secret or PII pre-filter. The README's recommended install is a prompt pasted into an agent plus a downloaded shell script.

Observed

License
AGPL-3.0-or-later (pyproject.toml, package.json)
Language
Python 3.12 to 3.14, with an npm wrapper requiring Node 18+
Packaging
PyPI and npm packages, both exposing the slm command; uv tool install also documented
Interfaces
CLI, MCP server (stdio and local HTTP), REST, web dashboard
MCP registry entry
server.json lists PyPI (uvx) and npm (npx) stdio packages
Platform support
Apple Silicon macOS, 64-bit Windows, 64-bit Linux; Intel Mac and 32-bit Windows unsupported
Declared maturity
Development Status 4 - Beta classifier in pyproject.toml
Storage
SQLite canonical store with LanceDB and CozoDB projections
Dependency policy
Runtime dependencies hard-pinned to exact versions, including torch and transformers
Default reranker
English-only cross-encoder; remote rerank endpoint configurable
Provider key storage
Plaintext in the config file with owner-only permissions, or environment variables
Support gate held-out results (docs)
77 of 77 answerable kept, 152 of 156 nonsense rejected, 7 of 50 unanswerable on-topic caught

Read from README.md, pyproject.toml, package.json, server.json, src/superlocalmemory/__init__.py, docs/ARCHITECTURE.md, docs/getting-started.md, docs/mcp-tools.md, docs/recall.md, docs/configuration.md.

Tags

agent-memorylocal-firstai-agentsplugingovernancememory

Tech Stack

Node.jsPython

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.