sqlite-graph-memory: persistent memory for Claude Code, Cursor and Codex over your own markdown notes
Long-term memory for coding agents that lives in the notes you already keep: 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.
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.
Three commands
pip install 'sqlite-graph-memory[embeddings]'
sgm-index ./notes
sgm-ask --ab "In which programming language is the tool behind Project Ouzel written?"
./notes is any folder of markdown notes. 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.
From any MCP client (Claude Desktop, Cursor, Zed, Windsurf): sgm-mcp after the install
above, or uvx --with 'sqlite-graph-memory[embeddings]' sqlite-graph-memory. Set
BRAIN_INDEX_DIR to the index folder sgm-index wrote. Config:
examples/mcp-config.json. Claude Code:
claude mcp add sqlite-graph-memory -e BRAIN_INDEX_DIR=/path/to/index -- sgm-mcp.
Codex: a [mcp_servers.sqlite-graph-memory] table in ~/.codex/config.toml with
command = "sgm-mcp". Agents that read files first: llms.txt.
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 a cross-encoder
reranks what the hop dragged in.
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]], andbrain_ask.pyjust 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 expensive part of RAG quality is a reranker, not graph infrastructure — see the
rerank step in
brain_ask.py.
So the pilot's bet: vector retrieval for entry points, hand-curated wikilinks for
association, a cross-encoder to keep serendipity honest, SQLite for everything that must
persist. All four 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. rerank cross-encoder → top-12
│ 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.
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 reranker 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_entityinbrain_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 inbrain_ask.pyis deliberately conservative, and switches the graph off only for name-shaped queries. - Crash-safety as a policy. The graph hop (
brain_ask.py) and the Stop-hook (turnstate_hook.py) are both wrapped, so any failure degrades to the previous behavior: vector-only recall, no ledger row. A memory layer must never make the agent worse than having no memory layer. - Measure, don't believe.
--abruns both pipelines on every real query and logs how many notes the graph promoted into the top-N, toab_recallinschema.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.
- No incremental indexing (tracked in the issues). 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.pyandmcp_server.pyhave unit tests, andturnstate_hook.pyhas write-path coverage. Runpytest -qfor 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 the embedding stack; the base install leaves it out
# so that `pip install sqlite-graph-memory` stays small instead of pulling ~2 GB
# of torch. Ask for it explicitly:
pip install 'sqlite-graph-memory[embeddings]'
# 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:
sgm-turnstate-show --stats
Working from a clone instead of an install? pip install -e '.[embeddings]' 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.
What it cost us to find out
Numbers from the home vault (~24k chunks, RU+EN, 2026-09-16), one frozen gold set of 184 questions (60 title / 60 body / 60 bridge / 4 temporal), vector+graph mode:
Those title and temporal nDCG figures are corrected, and the correction is the most
useful thing on this page. An external review panel reading this module found that ndcg_at
credited a gold note twice when the same filename appeared twice in the ranking — and since
wikilinks address notes by filename, that happens whenever two folders hold a foo.md.
Reproduced immediately: ndcg_at(['foo','foo'], ['foo']) returned 1.63, above the
maximum the metric can have. It affected 4 of our 184 questions, which inflated title nDCG by
0.031 and temporal by 0.108. Recall@12 and MRR were never affected (one is set-based, the
other stops at the first hit), and because the same four questions inflated both runs, every
before/after delta survived to the fourth decimal: title +0.0056 became +0.0055. The merge
and revert decisions below therefore stand — but the absolute levels we would have quoted
were wrong, and an eval nobody attacks keeps being wrong politely.
Two changes earned that. Cleaning duplicate snapshot copies out of the index (27 095 ->
24 216 chunks) moved the metric by exactly 0.0000 — it buys nothing on the ruler and stops
stale copies eating retrieval slots, which is worth knowing precisely because the number is
zero. The change that did move it gave the graph a vote in the ordering rather than in the
candidate set: personalized PageRank over the wikilink graph, fused with the reranker by RRF,
with the reranker's top 5 positions protected. The root cause was found by measuring instead
of guessing — 97% of the correct bridge answers were already in the candidate pool and the
cross-encoder was dropping them, because the relation is structural, not textual. Widening
retrieval, the obvious fix, would have done nothing at all.
That lane ships in the package as src/sqlite_graph_memory/ppr.py, off by default: turn it
on with SGM_PPR=1 or sgm-ask --ppr, and score it with python eval/run_eval.py --gold ... --ppr, which adds a third mode next to vector and graph. The numbers above were measured on
the graph pool, so the matching command is sgm-ask --graph --ppr; --ppr alone reorders the
vector pool, and --ab leaves the lane out so its A/B log stays vector against graph. Any failure in it falls back to
the reranker's order, like the graph hop. Be clear about what that reproduces: the bridge
numbers above came from the maintainer's home vault and home pipeline on 2026-09-16, and they
have not been re-run from the published package. The code is the same algorithm with the same
constants (alpha 0.85, 10 PPR files, top 5 protected), but your vault will give your numbers,
and bench/ does not score this lane yet.
Seven other changes were rejected by the same ruler, under one rule fixed in advance: a regression worse than 0.01 nDCG@12 on any class reverts the change.
None of those verdicts is a claim about the models. They are claims about this corpus with this ruler, which is the only kind of claim a swap decision actually needs — and the reason the module is here rather than in a blog post is that your corpus will disagree with ours.
What this eval cannot tell you
Wikilink gold is weak supervision: a linked neighbour is not always the note that answers the
question. bridge is built from links, so any method that walks links is flattered there —
it measures "can the pipeline follow a connection", not "is this the right connection". One
run scores one version of one index, and rebuilding the gold set between two runs makes the
two numbers incomparable, which is the quiet way an eval starts lying. On 60 questions a
±0.02 nDCG difference is inside the noise: we learned that by watching two candidate freezes
of the same pipeline disagree on 33 questions out of 184.
The metric can be broken on purpose. EVAL_MUTANT=1 makes nDCG return a perfect 1.000 for
everything, and pytest tests/test_eval.py must go red under it. A suite that stays green
while the metric is broken never tested the metric.
Roadmap
Now — latest release (v0.3.1 on PyPI and in the MCP Registry as of 2026-10-19).
The pilot as it runs daily: retrieval pipeline, per-turn ledger and A/B telemetry, all in
brain_ask.py and turnstate_hook.py, plus the bi-temporal design note. Known defects live in
the tracker rather than in footnotes
here, and this file deliberately names no issue numbers: they get closed, the prose around them
does not notice, and that drift is the defect this paragraph used to carry twice over.
Next:
- The public benchmark now exists as
bench/(bench-v0.1): a synthetic 202-note vault, 91 questions, a leaderboard that other systems can join, and CI that reproduces the numbers. Still missing: the full ablation matrix (hops × seed caps × neighbour caps × rerank pool × gating policy).bench/has its first two cells, 0 hops and 1 hop. The interesting question was never "does graph help" but for which query classes;eval/answers that per class instead of per intuition. - Bi-temporal edges — design note
docs/bitemporal.md. When you materialize the graph instead of parsing it at query time, give every edge a validity window (valid_from/valid_to/observed_at) so a rebuild closes superseded facts instead of deleting them: history is kept, recall prefers the present, andas_ofqueries can reconstruct the past. Non-destructive aging by confidence decay. The day-1 A/B result — context size ~unchanged, retrieved candidates ~60% fresher by mean age — is written up with its caveats indocs/bitemporal.md: the volume win is longitudinal, so that is one day on one corpus, not a benchmark. - A write-up on the pattern ("Graph RAG without graph extraction") is in progress.
Every noticeable change ships as a new release, so the release feed — not the commit graph — is where you can see whether "pilot" has stopped being the right word.
Models
- Embeddings:
intfloat/multilingual-e5-base(multilingual; the home corpus is RU+EN) - Reranker:
cross-encoder/mmarco-mMiniLMv2-L12-H384-v1
Both are set as constants (E5_MODEL, RERANK_MODEL) in brain_ask.py and small enough
to run on a laptop GPU; swap freely, since nothing in the pipeline depends on these
specific models.
Maintainers
- Anton Dziatkovskii (@tonydzi): retrieval pipeline, ledger, bench, releases.
- Vardhman Gupta (@Kaap10): the MCP server (
sgm-mcp, PRs #8, #10, #14) and its tests; reviews onmcp_server.pyroute to him via CODEOWNERS.
AI contributors
This project is built by a human + AI team, and the git log says so: Claude writes most of the code, Codex and Grok review it, Gemini feeds the research. Each is credited on a commit only if its output changed that commit's content — no decorative credits. Lab-wide policy, one source for every repo: AI-CONTRIBUTORS.md.
Cite this work
The paper about this system — retrieval design, pilot telemetry, and the query-class-stratified evaluation protocol — is published: DOI 10.5281/zenodo.22639718 (preprint, CC BY 4.0). Machine-readable citation: CITATION.cff (GitHub's "Cite this repository" button). Academic identity: Anton Dzyatkovsky publishes as Anton Dziatkovskii (ORCID 0000-0001-7408-3054).
License
MIT
Contact
Questions, war stories, or you want to run this on your own fleet:
- 👤 Author: Anton Dziatkovskii — Telegram @tonydzi · WhatsApp +1 341 222 9178 · X @Tony_Stef_
- 📣 Channels: @ClawRus (RU) · @ClawEng (EN)
- 🌐 palo-alto.ai · Palo Alto AI Research Lab
- 🧪 Engineers: want to test-drive this setup? Message me — I hand out free starter seeds to engineers who test and report back.
Tell us what is missing
You tried this on your own notes and something was missing or broke. That is the most useful thing anyone can tell us.
- You started and it did not work: open the 3-question form (what you tried, what broke, which machine).
- Not sure it is a bug, or just an idea: Discussions → Ideas.
Nothing in this package phones home. There is no telemetry, so we only know what you write to us.
Contributors welcome — and there is a queue
Issues labelled accepted
are scoped, free to take, and nobody is on them. Comment "claiming this" — no permission needed —
and it is yours for 7 days. New here? Start with
good first issue.
You keep the copyright to your code. No CLA, no assignment, ever — your contribution goes in under this repo's existing license, the same terms as ours. We answer every issue and PR within 48 hours, including "no, and here is why"; our silence is our bug, so ping the thread.
Full deal: CONTRIBUTING.md
🧩 One piece of a working system
This repository is one piece lifted out of a live operation: one engineer running a fleet of machines, built with Claude as implementation collaborator; the machines reach consensus with each other and wake the human only for money or the irreversible. It was extracted after it survived production, not written as a demo — and it runs on its own: nothing here phones home to the rest.
See how the whole thing fits together → SYSTEM.md
Its closest neighbours in the memory layer: second-brain-starter-kit · voice2brain
READ THIS WITH AI
One click and an agent reads the repo, pulls out the patterns and helps you apply them to your own work.
Read this repo: https://github.com/tonydzi/sqlite-graph-memory (“sqlite-graph-memory” - Graph RAG on SQLite for AI agents: vector retrieval + hand-curated wikilink graph + cross-encoder rerank, with a zero-token per-turn memory ledger. Working pilot). Work out what problem it actually solves, pull out the reusable patterns and help me apply them to my own setup. Start by asking what I am working on.
— TonyDzi, Palo Alto AI Research Lab · second brain, agent coordination, persistent memory: github.com/tonydzi
