sqlite-graph-memory MCPサーバー

sqlite-graph-memory: persistent memory for Claude Code, Cursor and Codex over your own markdown notes

Your coding agent remembers what you already wrote down, and reaches the note a plain search misses by following the links between your notes. It is an MCP server and a CLI over any folder of markdown, Obsidian vaults included: pip install, point it at the folder, and Claude Code, Cursor, Codex or any other MCP client can recall from it.

It fits Andrej Karpathy's LLM Wiki pattern (April 2026), where an agent maintains an interlinked markdown wiki and a search tool over the pages is the first helper he suggests adding: sgm is that search layer, read-only, and it follows your [[wikilinks]] as a graph. Independent project, not affiliated with or endorsed by him.

Your agent forgets everything between sessions, and plain vector search gives it back flat, unrelated chunks. This gives it associative recall over your own markdown notes: vector search finds the entry points, the [[wikilinks]] you already wrote are the graph. One SQLite file, no graph database, no ETL pass, no server.

Install in your client

Already using your own agent rules and projects? The Russian existing-setup guide combines this memory backend with the optional starter-kit method, preserves existing settings, and checks recall in a fresh chat. Install one memory connection per client; the starter-kit memory plugin and manual registration are alternatives.

Build the index once (any folder of markdown; uv gives you uvx):

BRAIN_INDEX_DIR=$HOME/.sgm-index uvx --python 3.12 --from 'sqlite-graph-memory[lite]' sgm-index ~/notes

--python 3.12 is deliberate. onnxruntime has no wheels for Python 3.14 on an Intel Mac, or on Apple Silicon below macOS 14. On such a machine an unpinned uvx falls back to release 0.3.5, which has no [lite], and sgm-index then stops and asks for [embeddings].

Then point your client at the server. Use the absolute path of that index folder.

Cold-start timeout (0.3.5 and later). The MCP server starts recall in a separate process with a default deadline of 180 seconds; a first call on a clean Windows install took 89 s, which the old fixed 60 s always cut off. Override with BRAIN_RECALL_TIMEOUT (integer seconds, 1 to 600; invalid values return an error). Your MCP client's own tool timeout must also allow the longer call. On 0.3.3 and earlier the deadline is a fixed 60 s: upgrade, or use sgm-ask --ask directly, and do not report MCP recall as verified merely because the CLI works.

Claude Code

claude mcp add sqlite-graph-memory -e BRAIN_INDEX_DIR=$HOME/.sgm-index -- uvx --python 3.12 --with 'sqlite-graph-memory[lite]' sqlite-graph-memory

Cursor (~/.cursor/mcp.json; Claude Desktop, Zed and Windsurf take the same mcpServers block)

{
  "mcpServers": {
    "sqlite-graph-memory": {
      "command": "uvx",
      "args": ["--python", "3.12", "--with", "sqlite-graph-memory[lite]", "sqlite-graph-memory"],
      "env": { "BRAIN_INDEX_DIR": "/absolute/path/to/.sgm-index" }
    }
  }
}

VS Code (.vscode/mcp.json in the workspace)

{
  "servers": {
    "sqlite-graph-memory": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--python", "3.12", "--with", "sqlite-graph-memory[lite]", "sqlite-graph-memory"],
      "env": { "BRAIN_INDEX_DIR": "/absolute/path/to/.sgm-index" }
    }
  }
}

Codex (~/.codex/config.toml)

[mcp_servers.sqlite-graph-memory]
command = "uvx"
args = ["--python", "3.12", "--with", "sqlite-graph-memory[lite]", "sqlite-graph-memory"]
env = { BRAIN_INDEX_DIR = "/absolute/path/to/.sgm-index" }

Size, du -sh of a fresh Python 3.12 venv on an x86_64 Mac (1 Nov 2026): the base install is 26 MB and has no torch (numpy only, enough for the graph half and the SQLite ledger); with [embeddings] (torch) it is 879 MB, 540 MB of that torch 2.2.2 CPU. With [lite], the same two models on ONNX Runtime and no torch, it is 155 MB (Windows, Python 3.12, 3 Nov 2026), installs in 12 s, and one MCP recall on the 202-note bench vault took 7.0 s against 33.8 s through torch on the same machine, with the same top hits and scores. [lite] is picked automatically when installed; SGM_BACKEND=torch forces torch (for a GPU). An index built by either backend is searched correctly by the other. Already installed with pip? Replace uvx --python 3.12 --with ... sqlite-graph-memory by sgm-mcp (examples/mcp-config.json). Agents that read files first: llms.txt.

Three commands

pip install 'sqlite-graph-memory[lite]'
sgm-index ./notes
sgm-ask --ab "In which programming language is the tool behind Project Ouzel written?"

On an Intel Mac, use Python 3.12 with pip install. On a newer Python it can quietly install 0.3.5, which has no [lite].

./notes is any folder of markdown notes. Re-running sgm-index only re-embeds notes that are new or changed since the last run (tracked by a manifest next to the index); pass --full to force a complete rebuild.

To get exactly the output below, use the 202-note synthetic vault from this repo (it is not inside the pip package):

git clone --depth 1 https://github.com/tonydzi/sqlite-graph-memory && cp -R sqlite-graph-memory/bench/vault ./notes

Real output, top 5 of 12:

--- DIRECT memory (vector) ---
 1. [  1.66] project-ouzel
 2. [ -1.33] project-avocet
 3. [ -1.79] tool-cobbleset
 4. [ -2.44] meeting-2025-10-21-lighthouse-sync
 5. [ -2.70] person-severin-ibarrondo

--- ASSOCIATIVE memory (vector+graph) ---
 1. [  1.66] project-ouzel
 2. [  1.14] tool-lanternfish  <== +graph
 3. [ -1.33] project-avocet
 4. [ -1.79] tool-cobbleset
 5. [ -2.44] meeting-2025-10-21-lighthouse-sync

The answer lives in tool-lanternfish ("Lanternfish is written in Elixir"). The question never names Lanternfish, so the vector pass does not find it in the top 12 at all. project-ouzel links to it, so one graph hop puts it second.

If that hop found something your vector search missed, a star is how other people with a notes folder find this repo. sgm-ask --ab asks the same once per machine, in a terminal only; SGM_NO_STAR_PROMPT=1 turns it off.

That is the idea in one query, and it is not a blanket win. On this same vault vector-only ranks ahead of graph mode overall; the numbers are in bench/. --ab exists so you can measure which one wins on your own notes.

What it is

Graph RAG on SQLite for AI agents — a working pilot, not a framework. It is the extracted memory layer of a personal "second brain" agent setup: vector search finds the entry points, the [[wikilinks]] you already wrote by hand are the graph, and dense similarity orders what the hop dragged in (a cross-encoder rerank is opt-in with --rerank; see below for why it is no longer the default). SQLite is the only database (schema.sql).

Status: pilot, and the word is load-bearing: it runs daily in one real setup (an index of ~28k curated notes, ~93k chunks, rebuilt nightly from a ~270k-file Obsidian vault driven by Claude Code), but it is deliberately minimal and makes no attempt to be general — see What's intentionally missing, which is where the uneven test coverage is described. Published as the companion code for an upcoming write-up on lightweight Graph RAG for agents; the one design note that is already written is docs/bitemporal.md.

Why

Most Graph RAG stacks assume a graph database, an ETL pipeline, and an entity-extraction pass — HippoRAG (arXiv:2405.14831) derives its KG that way, and Graphiti/Zep maintain a temporal one. For a single-user agent over a markdown knowledge base, all three are overkill:

• The graph already exists — Obsidian/Logseq/plain-markdown users hand-curate their edges as [[wikilinks]], and brain_ask.py just parses them. No extraction pass, and edge quality beats anything a model would mine.
• SQLite is enough — the only things worth persisting are derived state (a per-turn ledger) and telemetry, and both are defined in schema.sql. One file, stdlib driver, zero ops. The ledger is telemetry, not memory: recall never reads it, so past turns do not change what brain_ask.py or the MCP server return.
• Measure before adding a stage. The pilot shipped a cross-encoder reranker from day one; measured later, it lowered recall on long queries (our 184-question gold set: body nDCG@12 0.640 reranked vs 0.815 dense; DolphinBench Riley slice, 1000 sessions: R@5 0.48 vs 0.61). Since 0.3.7 it is off by default and stays available as --rerank / SGM_RERANK=1.

So the pilot's bet: vector retrieval for entry points, hand-curated wikilinks for association, SQLite for everything that must persist. All steps live in one file, brain_ask.py, in that order.

Architecture

Three layers, three failure-isolated components:

markdown notes (with [[wikilinks]])
        │
        ├── index_notes.py      chunk + embed (e5-base) → .npy/.pkl index
        │
        ├── brain_ask.py        recall pipeline:
        │       1. dense retrieve  top-60 chunks, dedup by file
        │       2. --graph         +1-hop wikilink neighbours of top-15 hits (max 40)
        │       3. order           dense similarity → top-12
        │                          (--rerank: cross-encoder instead, opt-in)
        │       4. --ab            run vector-only AND vector+graph, diff,
        │                          log the delta to SQLite (ab_recall)
        │
        └── turnstate_hook.py   agent Stop-hook: after every assistant turn,
                                parse the transcript tail and append one row
                                (ask, summary, files, tools, commands, decisions)
                                to SQLite (turns). 0 LLM tokens, pure stdlib.
                                turnstate_show.py = read-only viewer.
                                Off the recall path: nothing in recall
                                reads this table.

schema.sql documents both tables. The graph is not materialized in SQL: brain_ask.py parses edges from the notes at query time, bounded to 1 hop / 40 neighbours. That keeps the graph permanently in sync with the notes at zero maintenance cost; materialize an edge table only when hop depth or corpus size demands it.

Design choices that survived contact with reality

• Graph expansion is candidate generation, not ranking (brain_ask.py). Neighbours go into the same ranking pool as vector hits, so an irrelevant linked note gets buried. That is what makes 1-hop expansion safe to leave on.
• Entity gate (looks_like_entity in brain_ask.py). A/B telemetry showed graph expansion helps theme queries ("approaches to agent memory") but hurts name/tool lookups ("NotebookLM"): a name's card links to everything, so the hop pulls in noise. The heuristic in brain_ask.py is deliberately conservative, and switches the graph off only for name-shaped queries.
• Crash-safety as a policy. The graph hop (brain_ask.py) is wrapped, so a failure degrades recall to vector-only. The Stop-hook (turnstate_hook.py) is wrapped too, but it sits off the recall path: a failure costs one ledger row and nothing else. A memory layer must never make the agent worse than having no memory layer.
• Measure, don't believe. --ab runs both pipelines on every real query and logs how many notes the graph promoted into the top-N, to ab_recall in schema.sql. Whether "associative memory" earns its keep is a weekly SQL query, not a vibe.

What's intentionally missing

• No entity lane. The production setup has an extra retrieval lane over people/project cards; it is too entangled with personal data to publish.
• Packaging exists since v0.3.0 (pip install sqlite-graph-memory). There IS an eval suite now — see Measure before you swap — but it scores retrieval on wikilink-derived gold, which is weak supervision and says so out loud. Test coverage is uneven: index_notes.py, brain_ask.py and mcp_server.py have unit tests, and turnstate_hook.py has write-path coverage. Run pytest -q for the current count rather than trusting a number written in prose here — that is how this file ended up claiming "no tests" while seven were passing. The status line above says pilot on purpose.

Quickstart

# Indexing and reranking need an embedding stack; the base install leaves it out.
# [lite] runs the models on ONNX Runtime without torch (~155 MB); [embeddings]
# brings torch (~2 GB) if you want a GPU. Ask for one explicitly:
pip install 'sqlite-graph-memory[lite]'

# 1. index a folder of markdown notes
sgm-index /path/to/notes

# 2. ask, three ways
sgm-ask "how do I think about agent memory"
sgm-ask --graph "how do I think about agent memory"
sgm-ask --ab    "how do I think about agent memory"   # logs the diff to SQLite

# 3. inspect the A/B telemetry
sqlite3 turnstate.db "select ts, query, new_in_top, promoted_via_graph from ab_recall order by id desc limit 10"

Run any of them on a base install and they tell you which extra is missing rather than raising ModuleNotFoundError — that path is covered by tests/test_optional_extra.py.

To enable the per-turn ledger in Claude Code, register the Stop hook (the exact block is examples/claude-code-stop-hook.json, which calls sgm-turnstate-hook), then read it back with sgm-turnstate-show. The ledger records what each turn did; recall never reads it.

sgm-turnstate-show --stats

Working from a clone instead of an install? pip install -e '.[lite]' gives you the same four commands, and the modules stay importable as from sqlite_graph_memory import brain_ask.

Configuration is env-only and every variable is listed in .env.example; everything defaults to the current directory. GPU is used automatically if a CUDA torch build is present, and CPU works fine for small corpora.

What --ab looks like

Output of brain_ask.py with --ab, on a toy 4-note corpus (real runs use an index of ~28k notes / ~93k chunks):

A/B RECALL: how should agent memory work
(Direct=4 vector / Associative=+0 graph; Associative promoted 0 new notes into top-12, 0 of them via graph)

--- DIRECT memory (vector) ---
 1. [  4.11] agent-memory
 2. [ -6.50] graph-rag
 3. [ -8.66] sqlite-ledger
 4. [ -9.73] cooking

--- ASSOCIATIVE memory (vector+graph) ---
 1. [  4.11] agent-memory
 2. [ -6.50] graph-rag
 3. [ -8.66] sqlite-ledger
 4. [ -9.73] cooking

On a corpus this small the graph adds nothing — every note is already in the candidate pool. The interesting deltas appear at scale, and that is exactly what the ab_recall table in schema.sql accumulates evidence for.

Public benchmark — bench/

Every number in the next section comes from our private vault, so nobody outside the lab can check it. bench/ is the reproducible half: a generated 202-note vault about a fictional lab, 91 questions in seven classes (including 22 whose answer is not in the vault), and three metrics that matter for agent memory: recall@k on the support notes, verbatim citation, and abstaining when the fact is missing.

python bench/run.py --system bm25          # one command, stdlib only
python bench/validate.py                   # what CI runs on every result file

Our own baseline is on the leaderboard with its weak spots written out: on "what is the current deadline" questions it retrieves both decisions (R@5 1.00) and cites the superseded one 6 times out of 6; vector-only ranks ahead of graph mode on this vault (R@5 0.88 vs 0.85). Other memory systems (mem0, Letta, Khoj, plain RAG) can join the board with one adapter file and a pull request: bench/CONTRIBUTING.md. CI recomputes every submitted number from the per-question outputs, so a typed-in number cannot get onto the board.

Health check — sgm-stats

Four numbers about the link graph itself, no model and no index needed. It reads the notes with the same filter and the same [[wikilink]] parser as recall, so it measures the graph --graph actually walks. Thresholds follow second-brain-os metrics; each is a --threshold-* flag. Exit code 0 ok, 1 warn, 2 fail, so it can sit in cron.

$ sgm-stats bench/vault --today 2026-10-20
202 pages in bench/vault

FAIL  orphan_rate       26.2%   53 of 202 pages have no inbound link
OK    avg_degree         4.65   940 resolved links over 202 pages
OK    main_component   100.0%   202 of 202 pages; 1 component
WARN  stale_concepts   100.0%   12 of 12 concept pages older than 90 days (as of 2026-10-20)

The synthetic bench vault fails the orphan check for a plain reason: its 40 meeting and 12 decision notes link out and nothing links back to them. --json gives the same report with the first orphan and stale page names.

Measure before you swap — eval/

Every "obvious" upgrade to a RAG pipeline — a newer embedder, a stronger reranker, hybrid BM25, a different chunker — is a hypothesis. Without a number on your corpus, swapping one in is superstition with a changelog entry. eval/ is the smallest honest ruler we could build, and it needs no labelling budget and no LLM calls: the wikilinks you already wrote are the relevance labels.

python eval/run_eval.py --selftest                       # metrics only: no model, no index
python eval/build_gold.py /path/to/notes --n-per-class 60   # -> eval/gold-<today>.jsonl (freeze it)
python eval/run_eval.py --gold eval/gold-<today>.jsonl --sqlite

Four question classes, each probing a different part of the pipeline: title (a note's own title should find it), body (its first paragraph should find it), bridge (its first paragraph should surface the notes it [[links]] to — this is what the graph is for), and temporal (an outdated note's title should surface its replacement, via a superseded_by: "[[newer-note]]" frontmatter field). Every question is scored in both modes, vector-only and vector+graph, so "does the graph help?" stops being a matter of taste.

run_eval.py imports brain_ask.py and calls the same functions the agent calls — load_index, retrieve_candidates, expand_1hop, rerank_candidates — so what you measure is what runs. An eval that re-implements the pipeline measures the re-implementation. --sqlite appends one row per class and mode to a gold_eval table, which is how a run six weeks from now can be diffed against today's.

Sweeping the graph-expansion knobs — eval/run_ablation.py

run_eval.py scores one configuration at a time (vector vs graph). run_ablation.py sweeps a grid of pipeline parameters — whether the 1-hop graph expansion runs at all (hops), how many top hits seed it (ghops), how many neighbours it admits (gmax), the dense-retrieve pool size (topk), and whether entity/name-shaped queries skip the hop (entity_gate) — against the same gold file, in one run, reusing brain_ask.py and run_eval.py's own metrics so the sweep can't drift from what sgm-ask actually does:

python eval/run_ablation.py --list-grid                        # cells only, no model/index load
python eval/run_ablation.py --gold eval/gold-<today>.jsonl      # the built-in 10-cell default grid

Each dense embedding is computed once per question and reused across every cell (a cell's candidate pool is a slice of the widest pool asked for, not a re-retrieve); only the cross-encoder rerank runs per (question, cell), since the candidate set changes per cell. The report — printed as a Markdown table and written as both <prefix>.md and <prefix>.csv (--out <prefix>, or --csv/--md directly) — breaks down n, Recall@12, MRR, nDCG@12 and delta_nDCG@12 (vs the grid's 0-hop baseline) per (cell, class) and overall, and warns explicitly whenever a class has fewer than 20 questions. The built-in default 10-cell grid varies one axis at a time against 1-hop (ghops=15, gmax=40, topk=60, entity_gate=on) rather than a full Cartesian product; --grid takes default, a ';'-separated key=val cell list (e.g. --grid "hops=0;hops=1,ghops=30,gmax=80" to test multi-axis interactions), or a JSON array of cell dicts.

Sample run (intfloat/multilingual-e5-base + cross-encoder/mmarco-mMiniLMv2-L12-H384-v1 on CPU) over a 5-note synthetic mini-vault (12 questions: 5 title, 5 body, 1 bridge, 1 temporal) with --grid "hops=0,topk=3;hops=1,ghops=2,gmax=5,topk=3,entity_gate=on;hops=1,ghops=2,gmax=5,topk=3,entity_gate=off":

cell	class	n	Recall@12	MRR	nDCG@12	delta_nDCG@12
cell-0 (hops=0, topk=3)	bridge	1	0.500	0.500	0.387	+0.000
cell-0 (hops=0, topk=3)	overall	12	0.958	0.903	0.907	+0.000
cell-1 (hops=1, ghops=2, gmax=5, topk=3, gate=on)	bridge	1	1.000	0.500	0.693	+0.307
cell-1 (hops=1, ghops=2, gmax=5, topk=3, gate=on)	overall	12	1.000	0.903	0.933	+0.026
cell-2 (hops=1, ghops=2, gmax=5, topk=3, gate=off)	bridge	1	1.000	0.500	0.693	+0.307
cell-2 (hops=1, ghops=2, gmax=5, topk=3, gate=off)	overall	12	1.000	0.903	0.933	+0.026

When topk is smaller than the vault (topk=3 < 5 notes), 1-hop [[wikilink]] expansion (cell-1, cell-2) lifts bridge Recall@12 from 0.500 to 1.000 (+0.307 nDCG@12, +0.026 overall) without regressing title, body, or temporal. On the default 10-cell grid (topk=20..80 > 5 notes), every cell scores identically (overall Recall@12 1.000, nDCG@12 0.929, delta_nDCG@12 = +0.000) because dense retrieval already admits the entire mini-vault before the graph hop runs.