what is homestead-memory?

An open-source (MIT), local-first memory layer for AI agents. Your memory lives as plain markdown files on your disk, retrieval runs locally, and a verification gate (hsm verify) scores the memory's integrity 0-100 and fails loudly when it finds rot: contradictions, dead citations, silent drift. It works with Claude Code, Cursor, and Claude Desktop over MCP, and with anything else over a local HTTP API.

how is this different from Mem0, Zep, or MemPalace?

Two ways. First, it stores your memory verbatim in files you own instead of extracting lossy summaries at write time, so writes cost $0 and nothing is lost. Second, and nobody else does this: it verifies itself. Every extracted claim must cite a real source (checked mechanically, not trusted), and the whole store gets an integrity score we publish with every benchmark run.

are the benchmark numbers real?

Yes, and deliberately unglamorous. On the full 500-question LongMemEval set with the official judging methodology we score 85% retrieval recall and 52.8% QA accuracy, with every run, judge, and failed experiment published in the repo. Others self-report higher numbers on harnesses you can't run. Ours you can run tonight.

your ai's memory is either proven or prayed for

you’ve been renting your mind

every chat you have with a cloud ai teaches it something about you. your projects, your allergies, your deals, your drafts. that memory lives on their servers, under their terms, behind their subscription. cancel, and it was never yours. that’s not a memory. that’s a landlord with a filing cabinet.

so you run local ai. good. but here’s the part almost nobody talks about: local memory that can’t prove itself is just a smaller landlord. the file is on your disk, sure. is it right? is the fact your agent just confidently told you still true, or did it stop being true in march?

memory rots

memory doesn’t fail loudly. it rots quietly:

a note ends up contradicting itself and nobody notices
an extracted “fact” loses the source that justified it
a current value gets shadowed by a stale copy
the record drifts while its changelog moves on without it

every memory system we tested stores. none of them verify. they all ask you to hope.

proven, not prayed for

homestead-memory is our answer, and it’s open source (MIT). three commitments:

markdown you own. the memory is plain files. open them in any editor, git diff them, walk away with the folder any time. indexes are derived and disposable.

verification over trust. hsm verify runs mechanical integrity checks over the whole store and produces a number: 0 to 100, which we call rotbench. it exits nonzero on rot, so it gates your ci and your cron exactly like a test suite. no llm grading its own homework. a clean vault says MEMORY INTACT. a rotten one gets told, loudly, with file and line.

extraction you can audit. the optional distilled layer pulls entity facts out of your raw notes with verbatim quotes, and the quotes are checked in code against the source. a claim either cites something real or it gets dropped. when a fact changes, the old value doesn’t vanish: a changelog line records the transition with its source. extraction camp tools (mem0, zep) build structure you have to trust. ours you can check.

and it compounds, locally

a memory that just sits there is a smaller landlord too. hsm tune points at your own golden queries, measures which retrieval settings actually surface them on your vault, and keeps the best. the gain is a number you can see, the tuning only changes retrieval and never your notes (so hsm verify still gates it), and none of it leaves your machine. the frontier-lab pitch is “your usage trains our model.” this is the same loop, kept yours.

honest numbers, on purpose

on the full 500-question longmemeval benchmark (48-session haystacks, official judging methodology): 85% retrieval recall, 52.8% qa accuracy, ~5.2k context tokens per query, rotbench 99.4.

the qa number is honest and mid. competitors self-report way higher on harnesses you can’t reproduce. we publish the harness, the judge, the exact runs, and the experiments that failed. if you can beat our numbers with our own tooling, we’ll publish that too. the full run history is public, and rotbench is an open spec: break it and we merge your fixture.

sixty seconds to receipts

pip install homestead-memory

hsm init   ./my-vault
hsm ingest ./my-vault
hsm verify ./my-vault        # MEMORY INTACT, or you get told
hsm verify --demo            # watch it catch planted rot, live

claude code, cursor, or claude desktop:

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

this is the memory layer we described in cross-harness memory continuity, shipped as a real tool. it’s also the memory engine behind the upcoming homestead app: same files, same verification, with a MEMORY INTACT badge a non-technical person can read.

stop renting your mind. own it, and make it prove itself.

visit the lab →