the lab · open source · mit

stoprentingyourmind.

your ai's memory is either proven or it's prayed for. homestead-memory is a local, verifiable memory layer for ai agents: plain markdown you own, retrieval that works offline, and a gate that catches rot, tampering, and poisoning with mechanical checks, not vibes. every other memory tool asks you to hope it remembers. this one shows receipts. v0.2.2 isolates its qmd runtime, checks process ownership safely across macOS, Linux, and Windows, and reports every fallback instead of hiding it.

pip install homestead-memory

get it on github → read the story

install v0.2.2

sixty seconds. your machine.

Start with the CLI, verify the demo fixture, then connect the MCP server to the agent harness you already use.

the cli

pip install homestead-memory
npm install -g @tobilu/[email protected]

hsm init   ./my-vault
hsm ingest ./my-vault
hsm qmd start
hsm ask    "what did i decide about x?" \
  --retrieval balanced
hsm verify ./my-vault

claude code · cursor · desktop

claude mcp add homestead-memory \
  -- hsm mcp ~/my-vault

six mcp tools: ask, search, verify, history, ingest, distill. your agent gets a memory it can prove.

PyPI release → full docs →

watch it catch its own rot

memory intact. until it isn't.

hsm verify --demo

$ hsm verify --demo

building a clean vault: one fact, one linked note

MEMORY INTACT · 100/100 · 0 fail / 0 warn

planting rot: a note contradicts itself, a source is deleted,

an uncited claim slips into the distilled layer...

ROT DETECTED · 0/100

[self_contradiction] the note argues with itself

[dangling_citation] cited source no longer exists

[uncited_claim] distilled claim has no source

exit 1 · rot never ships

memory intact rot detected

that's the real cli. hsm verify exits nonzero on rot, so it gates your ci and your cron like a test suite. run it yourself: pip install homestead-memory

what "verifiable" actually means

memory fails three ways. it catches all three.

Rex the guard dog stamping a memory ledger INTACT 100 or ROT 0 — **rot.** a note contradicts itself, a source vanishes, a value goes stale. mechanical checks, scored 0 to 100.

Rex catching a note edited after it was signed, the wax seal cracking — **tamper.** a fact edited after it was signed. sign the vault and the ed25519 signature breaks. rex knows.

Rex catching a poisoned fake memory at the gate before it enters the store — **poison.** untrusted input slips in a memory that was never sourced. cite or it drops.

what it stands on

three commitments. zero black boxes.

markdown you own

the memory is plain files on your disk. open them, read them, git diff them, leave any time. indexes are derived and disposable. it's your folder.

verification over trust

memory rots quietly: notes contradict themselves, sources vanish, stale values shadow current ones. hsm verify scores integrity 0 to 100 and exits nonzero on rot. it gates ci like a test suite.

extraction you can audit

the optional distilled layer extracts facts with verbatim quotes, checked in code. a claim either cites a real source or gets dropped. contradictions append a changelog line, never a silent overwrite.

the culture is turning

we've been saying this for a month. now they are too.

the week we shipped this, the biggest podcast in tech spent thirty minutes on the same idea: stop handing your data and your edge to a frontier lab that can turn around and compete with you. all-in, ep 279.

"data retention is your treasure. transfer it at your own peril."

palantir manifesto, read on all-in #279

"why would you ever share proprietary data with them? you are mortgaging your future."

david sacks, all-in #279

they're describing the fortune-500 version: on-prem clusters, a server per employee, roll your own model. homestead-memory is the one you run on your laptop tonight. your memory, kept local, plain markdown you own, and it catches rot, tampering, and poisoning. free, and mit.

and the labs that host your model will happily ship the product that competes with you. ask figma. ask cursor. use claude code all you want. just don't hand it your memory.

honest numbers

benchmarked. receipts included.

cloud memory bills you per turn, forever. this is $0 to write, and it's a folder you already own.

99.4 / 100

rotbench integrity

the number that's actually ours. mechanical integrity checks over the store, no llm judging its own homework. recall is a crowded lane; nobody else scores whether the memory itself was tampered with or silently rewritten. self-scored on our vaults, so break it: the harness is open.

85%

retrieval recall@k

did the memory surface the right evidence? full 500-question LongMemEval set, 48-session haystacks with distractors. reader-independent.

52.8%

qa accuracy

honest and mid. others self-report higher on harnesses you can't run. ours you can, every failed run included.

~5.2k

context tokens / query

the cost axis. verbatim memory, $0 write-time cost.

every number above comes from a harness you can run yourself, judged with the official LongMemEval methodology. we publish our failures too: a lab whose valuation depends on looking infallible structurally can't do that. we can, so we do. the full run history is public.

break it. we dare you.

rotbench is an open spec

the integrity score is only credible if it survives adversaries. build a memory store that's obviously rotten but scores intact, or an intact one that false-positives, and we merge your fixture and fix the check. the scoreboard of merged breaks lives in the repo. trust you can inspect, and break, is the one thing a duopoly can't ship.

read the spec →