frugal-memory

A persistent memory for AI assistants that is just a SQLite file — and costs about a fifth of what reading markdown notes does.

No vector store, no embeddings, no external service, no API key. Python, SQLite with FTS5, stdio — the only dependency is the official mcp SDK. Nothing leaves your machine unless you explicitly turn on the optional fact check.

Measured over 275 real questions from 89 working sessions: 2,359 tokens per answered question, against 11,319 for the usual one markdown file per fact, found with grep and read whole. The gap grows as the memory does, because every answer here is capped and grep is not. All figures come from one installation, and the docs say where one did not hold up.

Why

An assistant's memory usually fails in one of two ways: it is a pile of markdown files that has to be read whole to be searched, or it is a vector database that answers that something matched but never why.

This one is a ranked full-text index over short entries, with two things bolted on that turn a note pile into a memory: entries are filed on two axes — where it belongs (project) and what kind of knowledge it is (pitfall, decision, measurement, …) — and when you write something down, the tool tells you which existing entries it overlaps with, so contradictions surface instead of quietly piling up.

Features

• Capped, calculable answers — a search returns eight preview lines of about 1.2 KB in total instead of 4.2 KB of full text; zeige then fetches the entries you pick, about 6,000 characters per call unless you ask for everything. The worst case is known before you ask.
• It learns from its own failures — when a search is followed by fetching an entry, that pair is remembered. Ask the same thing again and the entry that answered last time comes first. Word-for-word repeats only, and the header says when it fired.
• Two filing axes — a free-text project tag and a closed, enforced vocabulary of knowledge kinds. Chronicle entries are hidden from search by default, but the header always says how many were held back.
• Duplicate and contradiction hints — remember reports similar existing entries and which numbers changed, so you can supersede instead of accumulate.
• Nothing is ever deleted — superseded entries drop out of the default view and stay findable, with a note on why and by what.
• Browsing, not just searching — themen() lists projects and a project's newest title lines for when you have forgotten the words to search for. Capped like everything else: one page, and a line saying how to get the rest.
• Project-bound in use, not in search — the usual question is asked inside one project, with marke=. That filter is hard, so when the best match lies outside the project, recall names it instead of leaving you with a plausible-looking list of the wrong entries.
• Opt-in second opinion — optionally ask any OpenAI-compatible model whether two entries contradict each other or just supersede one another.
• German by default, English on request — tool names, messages and kinds are German; MEMORY_LANG=en switches all of it. The search also matches inflected forms and compound parts ("Rangliste" finds "Ranglisten"). Measured, that rarely changes the answer: the plain any-term stage of the search already finds what the stems would add (details).

Requirements

• Python 3.11+
• The mcp SDK (requirements.txt, that's the whole list)
• SQLite with FTS5 — included in standard CPython builds
• Optional, for the fact check only: any OpenAI-compatible endpoint

Setup

From PyPI, as a command in its own environment:

pipx install frugal-memory        # or: uv tool install frugal-memory

claude mcp add memory -s user -- frugal-memory

The database is created on first use at ~/.local/share/frugal-memory/memory.db ($XDG_DATA_HOME if set). frugal-memory-nachziehen is the upkeep script described in DESIGN.md.

Or from a checkout of this repository:

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

claude mcp add memory -s user -- \
    /path/to/frugal-memory/.venv/bin/python \
    /path/to/frugal-memory/memory_server.py

Here the database is created as memory.db next to the script. Either way, MEMORY_DB=/some/path.db puts it somewhere else.

-s user makes the server available in all projects; the entry ends up in ~/.claude.json. Restart Claude Code afterward — MCP servers are only loaded at startup. The tools are then called mcp__memory__recall and mcp__memory__remember.

For Claude Desktop, use ~/.config/Claude/claude_desktop_config.json instead:

{
  "mcpServers": {
    "memory": {
      "command": "/path/to/frugal-memory/.venv/bin/python",
      "args": ["/path/to/frugal-memory/memory_server.py"]
    }
  }
}

One rule matters more than any setting: one fact per call, not one document. recall returns whole entries — store a 180 KB file as a single entry and you get it back as a single hit, having gained nothing.

Teaching the agent how to use it

The tools alone don't tell an agent the working rules: search before reading files, recall → zeige, exactly one kind per entry, supersede instead of accumulate, retry in the other language. Those live in a skill that ships with the repo (not with the PyPI package):

cp -r skills/frugal-memory ~/.claude/skills/

Claude Code loads its full text only when the memory comes up; until then only the one-line description sits in context.

Reading another machine's memory

Each machine keeps its own memory, and each numbers its entries from 1. A session on machine A can read the memory of machine B over SSH — the server runs on B, next to its file; the database itself never crosses the network:

claude mcp add memory-b -s project -- \
    ssh b /path/to/frugal-memory/.venv/bin/python \
          /path/to/frugal-memory/memory_server.py --gast

--gast (guest) means three things:

• Read-only. The file is opened with mode=ro, and only recall, zeige and themen are offered. The writing tools are gone, and so are verdichten and pruefe: upkeep belongs to the owner, who can act on it. A guest's searches do not feed the owner's search feedback either, and a guest never migrates the schema.
• Every id carries the memory's name: b#72, not #72. The name is the host name in lower case; MEMORY_NAME overrides it.
• A named id is checked, in both modes. zeige("b#72") asked of any memory other than b is refused instead of quietly returning that memory's own #72. Bare ids (72, #72) mean the memory being asked.

-s project keeps the three extra tool definitions out of sessions that do not need them. The tools are then called mcp__memory-b__recall and so on.

The Eight Tools

remember(text, tags, art, ersetzt)	store an entry; reports similar existing ones and what it would supersede
recall(query, limit, art, marke, …)	ranked full-text search, preview by default; under marke= it names a better match outside the project
zeige(ids, voll)	full text of specific entries, about 6000 characters per call; what doesn't fit is listed with the call that fetches it
themen(marke, limit, seite, alle)	browse: which projects exist, or one project's title lines, newest first, a page at a time
vergessen(ids, grund)	mark as outdated — never deletes
einordnen(ids, art)	assign the knowledge kind of existing entries
verdichten(marke, art, …)	find groups of entries that say much the same thing
pruefe(ids)	opt-in: have a language model judge contradiction vs. newer state

Full reference with parameters and behavior: docs/TOOLS.md. With MEMORY_LANG=en the same tools come with English names and parameters, see A Note on Language.

Some Numbers

Measured on one real dataset that grew from ~900 to ~1,900 entries along the way — evidence from one installation, not a benchmark:

• 4.8× cheaper than reading markdown files, in real tokens, over 275 real questions (2,359 vs. 11,319 per question). The gap grows with the dataset and with file size, because reading a file is all-or-nothing: one question against a 188 KB project file costs 94,152 tokens.
• 89.8 % Recall@10 against the public LoCoMo dataset with a session as the unit, untuned (2026-10-19; the first run in September gave 82.3 %, and 45.1 % with a single dialogue turn as the unit).
• 73 % smaller returns from the preview, with a calculable worst case.
• −23 % expected characters per question from the feedback loop, at 84 % vs. 73 % first-round hit rate, measured over 164 questions from 30 real sessions in a time-ordered replay — a pair only ever helps later questions, never itself. 19 questions gained, none lost.
• 53 % → 64 % hit rate on 47 real reformulation chains from a one-line ranking fix — found by measuring where real search sessions had failed, not by guessing. On today's dataset the same fix also costs 3 of the 117 questions that already worked, so the net is +2 of 164.
• 93 % of the dataset belongs to a project (1,777 of 1,907 entries), and that is how the tool is used. Search itself does not care: general knowledge is found as well as project knowledge (92 % vs. 96 %, a tie under the 10 % rule). What the project filter buys is precision for short questions — 97 % vs. 85 % found at about three words. What it costs is everything outside the project: 40 questions about general knowledge, asked under 22 project tags, found 0 of 880. recall now names the better match outside the tag in 82 % of those cases, and in 70 % that line already is the entry sought; on real project questions it fires in 7–16 %.
• −83 % for a project's title index, −34 % for fetching eight full texts, from capping themen and zeige. On 110 project questions the cap cost not one answer.

Where a measurement did not survive a larger sample, it says so: docs/MEASUREMENTS.md, with the raw reports in messung/. The ranking figure above was corrected twice — see The Ranking Fix, Corrected Twice.

Documentation

docs/TOOLS.md	all eight tools in detail
docs/DESIGN.md	the two axes, the morphology, the database schema, and what was left out on purpose
docs/MEASUREMENTS.md	what was measured, and the known limitations
docs/FACTCHECK.md	the opt-in fact check: setup, what it can and cannot do
docs/IMPORT.md	importing an existing folder of notes, and the damage that did
messung/	the eight underlying measurement reports (mostly German)

A Note on Language

German is the default: tool names, status messages and the kind vocabulary (schnittstelle, fallstrick, entscheidung, messwert, arbeitsweise, verlauf, gemischt) are German, and the search stems German and splits German compounds.

English mode: start the server with MEMORY_LANG=en. Then

• the tools are remember(text, tags, kind, supersedes), recall(query, limit, kind, tag, full, include_superseded), show(ids), topics(tag, limit), forget(ids, reason), classify(ids, kind), consolidate(tag, kind, threshold, groups) and check(ids),
• the kinds are interface, pitfall, decision, measurement, workflow, history, unsorted,
• every message is English,
• the search uses a small English suffix stemmer ("configured" finds "configuring") instead of the German one, and does not split compounds.

The database itself does not change: kinds are always stored under their German names and translated at the border, and German kind names are accepted in both modes. What does change is the stem index, so the file records the language it was built in and the server refuses to open it with the other one. To switch an existing memory, back it up and run MEMORY_LANG=en python3 nachziehen.py once.

Not covered by the switch: the fact-check prompt stays German (it is calibrated word for word, and adding one sentence once halved its accuracy; only its verdicts are shown in English), and so do the import script and the measurement reports. The English stemmer is new and has not been measured on real questions the way the German search has.

Files

memory_server.py	MCP server, eight tools, search cascade, schema, migrations
morphologie.py	stemmer and compound splitter (German; small English stemmer for MEMORY_LANG=en)
werkzeuge_en.py	the English tool surface for MEMORY_LANG=en
texte_en.py	English messages, keyed by the German text
nachziehen.py	recomputes stems/teile and refreshes the tag directory
faktencheck.py	opt-in second opinion via any OpenAI-compatible model (off unless configured)
pruefstand_fc.py	measures the fact check against gold labels from the dataset
pruefpaare_gebaut.json	the hand-built and hand-found contradiction pairs
import_memos.py	import a file-based memory (one-off; see docs/IMPORT.md)
test_memory.py	regression net, one case per pitfall (no pytest)
skills/frugal-memory/SKILL.md	working rules for the agent, as a Claude Code skill
requirements.txt	just mcp
memory.db	the database — created on first use, never committed

messung/ keeps the eight measurement reports. The scripts that produced them and their raw data are not included: they only run against the author's own dataset, and the raw data is a verbatim transcript of real working sessions.

License

MIT — see LICENSE.