Potencialize seu agente
O comando do próprio publicador. Não é um endpoint do PlugMCP.
Host a copy with our keys and logs.
npx -y @jeffs-brain/[email protected]O Claude Code não tem link de instalação para um comando local. Copie o comando abaixo e execute-o em um terminal.
claude mcp add jeffs-brain-memory -- npx -y @jeffs-brain/memory@1.1.0Execute em um terminal.
{ "mcpServers": { "jeffs-brain-memory": { "command": "npx", "args": [ "-y", "@jeffs-brain/[email protected]" ] } }}claude_desktop_config.json
URL remota ou um comando na sua máquina? Como as duas instalações diferem
Rodar no PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
AbrirCross-language memory library for LLM agents. Three SDKs (TypeScript, Go, Python) that all implement the same spec so a brain created by one SDK reads and writes cleanly through another.
Local-first, hosted-optional. Apache-2.0.
All three SDKs implement the full spec wire surface. The shared HTTP conformance suite sits at 28/29 green across SDKs. Cross-SDK evaluation parity today is the shared daemon surface in exactly three scenarios: ask-basic, ask-augmented, and search-retrieve-only.
@jeffs-brain/memory-postgres) is TypeScript-only today. Go and Python use SQLite-backed stores.The shared runner covers those daemon scenarios only. It forwards retrieval mode unchanged and defaults to auto, so each SDK daemon resolves auto to hybrid when embeddings are available or bm25 otherwise. Full LongMemEval replay, replay ingest, and agentic loops stay in the native SDK runners; Go provides the reference memory eval lme run --ingest-mode=replay path. See eval/README.md for the shared verification flow and artefact layout.
These are the only cross-SDK daemon scenarios the shared harness verifies:
Parity expectation is explicit:
ask-basic and ask-augmented expose the same SSE contract: retrieve, answer_delta*, citation*, done.search-retrieve-only exposes the same JSON retrieval contract and is scored from returned retrieval content only.questionDate is forwarded for ask-augmented and search-retrieve-only when the dataset row provides it.candidateK and rerankTopN are forwarded on search-retrieve-only when the runner flags are non-zero.mode is forwarded unchanged by the runner; each daemon resolves auto locally.search-retrieve-only to replay memory only via actor-side retrieval filters: --actor-scope memory --actor-project <brain-id>. That keeps global memory plus the eval brain's project memory in scope while excluding raw transcript rows.ask-basic: the basic /ask request and SSE response path, including retrieve, answer_delta, citation, and done.ask-augmented: the augmented /ask prompt path, including readerMode=augmented, temporal anchor forwarding, and the same SSE contract.search-retrieve-only: the /search JSON contract only. The scorer consumes the returned retrieval content, not a daemon-generated answer.eval/runner.py starts one SDK daemon at a time, runs one scenario per invocation, and writes <output>/<YYYY-MM-DD>/<sdk>.json.eval/scripts/run_tri_lme.sh extracts once with Go, starts all three daemons against the same brain, and benchmarks search-retrieve-only only. Retrieval happens inside each daemon via /search; the shared augmented reader and judge stay in the Go runner process.Use a separate --output root per scenario when comparing SDKs on the same day. The runner writes one <sdk>.json per output root and date, so reusing the same output root for a second scenario overwrites the earlier JSON for that SDK.
ts, go, and py for the same scenario.eval/results/ for transport, citations, and answer or retrieval-blob shape.To compare one scenario across all three SDKs, run the shared runner from eval/ once per SDK:
cd eval
uv sync
# Fast offline parity check using a seeded reference brain
for sdk in ts go py; do
uv run python runner.py \
--sdk "$sdk" \
--dataset datasets/smoke.jsonl \
--scorer exact \
--scenario search-retrieve-only \
--mode bm25 \
--brain eval \
--seed-reference-brain \
--output results/smoke-search
done
# Corpus-grounded parity checks, using a populated `eval` brain
for sdk in ts go py; do
OPENAI_API_KEY=sk-... uv run python runner.py \
--sdk "$sdk" \
--dataset datasets/lme.jsonl \
--scorer judge \
--scenario ask-augmented \
--brain eval \
--output results/ask-augmented
done
for sdk in ts go py; do
OPENAI_API_KEY=sk-... uv run python runner.py \
--sdk "$sdk" \
--dataset datasets/lme.jsonl \
--scorer judge \
--scenario search-retrieve-only \
--brain eval \
--output results/search-retrieve-only
done
Compare eval/results/<scenario>/<YYYY-MM-DD>/ts.json, go.json, and py.json for the shared runner, or eval/results/tri-lme-<timestamp>/result-*.json for replay-backed tri-SDK retrieval parity.
Per-SDK regression commands for those scenarios live in the SDK READMEs below.
npm i -g @jeffs-brain/memory
memory init
memory ingest ./docs
memory search "question"
Full walkthrough: docs.jeffsbrain.com/getting-started/typescript/.
go install github.com/jeffs-brain/memory/go/cmd/[email protected]
memory version
memory serve --addr 127.0.0.1:18841
Full walkthrough: docs.jeffsbrain.com/getting-started/go/.
pip install jeffs-brain-memory
# or: uv add jeffs-brain-memory
memory init
memory ingest ./docs
memory search "question"
Full walkthrough: docs.jeffsbrain.com/getting-started/python/.
One command wires jeffs-brain into Claude Code, Claude Desktop, Cursor, Windsurf, and Zed:
npx @jeffs-brain/install
Or register a single host manually against any SDK's MCP wrapper:
# TypeScript
claude mcp add jeffs-brain -- npx -y @jeffs-brain/memory-mcp
# Go
claude mcp add jeffs-brain -- memory-mcp
# Python
claude mcp add jeffs-brain -- uvx jeffs-brain-memory-mcp
Each wrapper exposes the same 11 memory_* tools defined in spec/MCP-TOOLS.md.
Every SDK implements the full pipeline:
k=60.memory serve is exercised in ask-basic, ask-augmented, and search-retrieve-only.Provider contract, an in-process RBAC adapter (workspace -> brain -> collection -> document hierarchy with admin/writer/reader roles and deny:<role> overrides), a Store wrapper that runs a check on every read/write/delete, and a sibling OpenFGA HTTP adapter. The shared FGA model lives at spec/openfga/.spec/
openfga/ Canonical OpenFGA authorisation model shared by every SDK adapter
... Language-neutral protocol, storage, algorithms, query DSL, MCP tool contract
sdks/
ts/
memory/ Core TS SDK: store, retrieval, ingestion, query DSL, CLI, acl primitives
memory-postgres/ Postgres + pgvector adapter
memory-openfga/ OpenFGA authorisation adapter
go/
acl/ Authorisation contract, in-process RBAC, Store wrapper
aclopenfga/ OpenFGA HTTP adapter
... Full Go SDK: store/fs|git|mem|http, search, query, retrieval, memory, knowledge, eval/lme, cmd/memory, cmd/memory-mcp
py/
src/jeffs_brain_memory/
acl/ Authorisation contract, in-process RBAC, Store wrapper
acl_openfga/ OpenFGA HTTP adapter
... Full Python SDK: store, search, query, retrieval, rerank, memory, knowledge, ingest, llm, http daemon
mcp/
ts/ @jeffs-brain/memory-mcp (npm)
py/ jeffs-brain-memory-mcp (PyPI)
install/ @jeffs-brain/install, one-shot multi-agent installer
examples/ ts/hello-world and go/hello-world ready-to-run
docs/ Astro Starlight documentation site, Cloudflare Pages ready
eval/ Cross-SDK conformance and evaluation runner (--sdk ts|go|py)
The Go MCP wrapper ships from go/cmd/memory-mcp/. The mcp/go/ folder is reserved for future extraction.
Full docs site: docs.jeffsbrain.com.
Topic guides:
memory.config.json, env vars, CLI flagsspec/ - protocol, storage, algorithms, MCP tools, query DSL, conformance harnessdocs/ - public documentation site (Astro Starlight, Cloudflare Pages)examples/ - runnable hello-world per languageeval/ - cross-SDK evaluation runner, datasets, and verification workflowinstall/ - @jeffs-brain/install orchestratorCONTRIBUTING.md - dev setup, commit style, DCO, PR processSECURITY.md - vulnerability reportingCODE_OF_CONDUCT.md - community standardsNOTICE - bundled third-party componentsApache-2.0. See LICENSE.
Local-first memory and hybrid retrieval library for LLM agents with pluggable stores, BM25 plus vector search, and a four-stage memory pipeline.
{ "mcpServers": { "jeffs-brain-memory": { "command": "npx", "args": [ "-y", "@jeffs-brain/[email protected]" ] } }}Ferramentas
Ainda não há lista de ferramentas. Servidores remotos recebem uma após a primeira verificação de saúde.
Perguntas frequentes
Copie nesta página a configuração stdio para Cursor, Claude Desktop, Claude Code ou VS Code. Ela usa a instalação do publicador, não um proxy do PlugMCP.
Este registro não lista nenhuma variável de ambiente obrigatória.
@jeffs Brain/memory roda localmente via stdio.
Última verificação: 2026-11-15. Local-first memory and hybrid retrieval library for LLM agents with pluggable stores, BM25 plus vector search, and a four-stage memory pipeline.