Homestead Memory MCP Server

homestead-memory

Stop renting your mind.

Local-first, verifiable AI memory. Your notes stay plain markdown you can read, git diff, and own, and the memory catches its own rot, tampering, and poisoning.

Every other memory layer asks you to hope it remembers. This one lets you watch it catch the rot, live:

pip install homestead-memory      # Python 3.10+, macOS / Linux / Windows, zero deps
npm install -g @tobilu/qmd@2.1.0 # optional hybrid retrieval runtime

hsm verify --demo
# ① a clean vault           ✅  MEMORY INTACT — 100/100
# ② rot is planted…         🔴  ROT DETECTED —   0/100
#    🔴 [self_contradiction] the note argues with itself about its own status
#    🔴 [uncited_claim]      a distilled claim has no source citation
#    🔴 [dangling_citation]  a cited source no longer exists
#    ⚠️  [broken_link]        a reference points at a note that isn't there

Got No matching distribution found? macOS still ships Python 3.9 as its built-in python3, and this needs 3.10+. Nothing is wrong with the package. Either use a newer Python, or skip installing entirely: uvx --from homestead-memory hsm verify --demo (uv fetches a suitable Python for you.)

hsm verify exits non-zero on rot — it gates CI and cron like a test suite.

What did your agent actually do?

Agents fail quietly. The run reports success, the tool returns 200, and the thing you asked for did not happen. There is usually no record to contradict it.

hsm hook --install     # prints a Claude Code hook; you paste it, nothing is edited for you
hsm watch              # what your agent did, in order

#     0  14:02:11  Bash           npm test
#     1  14:02:19  Read           src/api/billing.py
#     2  14:02:24  Edit           src/api/billing.py

Every entry is hash-chained to the one before it, so editing, deleting, or reordering any record breaks every hash after it. hsm verify reports the break at the exact index. Sign it and a wholly rebuilt chain is caught too.

This is a file, not a platform. Agent observability tools are far richer than this and they want a deployment: the self-hosted ones document a production floor of several services and roughly 16 GB of RAM. This is pip install, one hook line, and a JSONL file on your disk. Different job. If you need dashboards, evals, and span analytics, use one of those. If you want a record you can grep and prove, use this.

Secret-shaped values are redacted and payloads truncated to a 200-character head, with a SHA-256 of the full original kept so the evidence survives redaction. That is a mitigation, not a guarantee: no pattern list is complete.

hsm export --evidence  # a pack anyone can verify with no install at all

The pack carries the records, the signature, the public key, an integrity report, and a standard-library verifier a third party can read in full and run. It states what it does NOT prove, including that a signature only establishes origin if you already know which key to expect.

Capture is Claude Code only right now. The MCP integration below works anywhere MCP does; the hook that records every tool call uses a Claude Code PostToolUse hook. Cursor and Codex need their own mechanisms and those are not built yet.

Quickstart (60 seconds)

hsm init   ./my-vault          # scaffold or adopt any markdown folder
hsm ingest ./my-vault          # index it (hybrid BM25+vector via qmd, optional)
hsm ask    "what did I decide about X?"
hsm ask    "what did I decide about X?" ./my-vault --budget 1200 --json
hsm search "what did I decide?" ./my-vault --retrieval balanced --json
hsm qmd start                  # persistent loopback runtime; no shared global index
hsm verify ./my-vault          # the integrity gate — the whole point
hsm distill ./my-vault         # optional: build the cited, verifiable fact layer
hsm history <note> --as-of 2026-06-01   # what was true THEN (temporal layer)
hsm serve                      # local HTTP API (auth'd, loopback-only)

Python agents can use the SDK directly:

from homestead_memory import connect

memory = connect("~/my-vault", agent="my-agent")
memory.remember("user", "city", "Berlin")
memory.ask("what city is the user in?")

The local HTTP API is documented in docs/openapi.yaml.

Retrieval profiles

Homestead keeps qmd in dedicated cache and config directories. It never runs maintenance against qmd's global index. Every structured result reports engine, retrieval_mode, degraded, reason, elapsed_ms, and index_age_seconds.

The route is persistent qmd MCP, then the dedicated qmd CLI, then a read-only direct scan. Run hsm qmd doctor, hsm qmd refresh, and hsm qmd status to inspect the runtime without touching any other qmd collection.

Refresh is explicit and incremental. It writes an atomic checkpoint beneath .hsm/refresh-state.json, refuses foreign or unhealthy QMD runtimes, emits a live heartbeat while QMD works, and commits the vault fingerprint only after embedding reaches zero pending vectors. Reads never trigger an implicit refresh; if QMD is unavailable, retrieval falls back to a read-only scan and reports the degraded engine and reason.

For a Linux/systemd reference deployment, see deploy/reference/.

Memory under the router

Routers can swap the served model while homestead-memory keeps the same vault underneath. The model name is just the runtime argument; provenance is stamped as name@model in the agent field when a write happens.

from homestead_memory import connect
from homestead_memory.adapters.openai_compat import MemoryChat

memory = connect("~/my-vault")

def remember_reply(response, memory, agent):
    memory.remember(
        "conversation",
        "last_reply",
        response.choices[0].message.content,
        source="chat",
        agent=agent,
    )

chat = MemoryChat(openai_compatible_client, memory, remember_fn=remember_reply)
chat.create(model="claude-sonnet-4.7", messages=[{"role": "user", "content": "brief me"}])
chat.create(model="glm-4.7", messages=[{"role": "user", "content": "continue"}])

memory.history("conversation")  # agents include assistant@claude-sonnet-4.7 and assistant@glm-4.7

LiteLLM can use the same pattern with a pre-call injection helper and a success logger:

from homestead_memory import connect
from homestead_memory.adapters.litellm_memory import MemoryLogger, inject_memory

memory = connect("~/my-vault")
messages = inject_memory([{"role": "user", "content": "brief me"}], memory)

# LiteLLM callback registration style depends on your app setup.
logger = MemoryLogger(memory, agent_name="assistant")

MCP already sits above harness-level routers. In a claude-code-router-style setup that swaps the backend model, homestead-memory keeps working with zero config because memory is external to the model. history() and verify then attribute every recorded fact to the exact name@model that wrote it.

Integrations

Adapters target the public framework interfaces listed here as of the current releases and may need version bumps as those APIs evolve. Core remains stdlib-only; install only the extra for the framework you use.

Universal tools work with any orchestrator that can register callables or JSON-schema function tools:

from homestead_memory import connect
from homestead_memory.adapters.tools import recall_tool, remember_tool, tool_specs, verify_tool

memory = connect("~/my-vault", agent="my-agent")
tools = [remember_tool(memory), recall_tool(memory), verify_tool(memory)]
specs = tool_specs(memory)  # name, description, parameters

LangGraph BaseStore (targets langgraph>=0.2):

from homestead_memory import connect
from homestead_memory.adapters.langgraph_store import HomesteadStore

store = HomesteadStore(connect("~/my-vault", agent="langgraph"))
graph = builder.compile(checkpointer=checkpointer, store=store)

CrewAI storage/memory (targets crewai>=0.70, storage-style save/search/reset):

from homestead_memory import connect
from homestead_memory.adapters.crewai_memory import HomesteadCrewAIStorage

storage = HomesteadCrewAIStorage(connect("~/my-vault", agent="crewai"))
storage.save("Researcher found the supplier shortlist", metadata={"task": "supplier_shortlist"})

AutoGen autogen_core Memory protocol (targets autogen-core>=0.4):

from autogen_core.memory import MemoryContent, MemoryMimeType
from homestead_memory import connect
from homestead_memory.adapters.autogen_memory import HomesteadAutoGenMemory

memory = HomesteadAutoGenMemory(connect("~/my-vault", agent="autogen"))
await memory.add(MemoryContent(content="Use metric units", mime_type=MemoryMimeType.TEXT))

OpenAI Agents SDK Session protocol or function tools (targets openai-agents>=0.0.1):

from homestead_memory import connect
from homestead_memory.adapters.openai_agents import HomesteadSession, function_tools

memory = connect("~/my-vault", agent="openai-agents")
session = HomesteadSession(memory, session_id="user-123")
agent_tools = function_tools(memory)

Claude Code / Desktop / Cursor (MCP: memory tools, not action capture):

claude mcp add homestead-memory -- hsm mcp ~/my-vault
# tools: memory_ask · memory_search · memory_verify · memory_history ·
#        memory_ingest · memory_distill

Why this exists

"Runs on your device" is table stakes now — every memory tool stores locally. Nobody verifies. Memory rots quietly: a note contradicts itself, an extracted "fact" loses its source, a body drifts past its own changelog, the current value gets shadowed by a stale one. You find out weeks later, when your agent confidently tells you something that stopped being true in March.

And rot is only the passive failure. Memory also gets tampered with (a fact edited after it was written) and poisoned (untrusted input injects a "memory" that was never true — a named 2026 attack class). homestead-memory catches all three mechanically: sign the vault and any edited byte breaks the signature; a distilled claim must cite a source that resolves or it's dropped. Recall benchmarks measure whether the model remembers; RotBench measures whether the memory can be trusted — see benchmarks/ROTBENCH.md.

homestead-memory is built around three commitments:

1. Markdown-primary. The human-readable files ARE the memory. Indexes and projections are derived and disposable. You can leave any time — it's your folder. Import/export Google's Open Knowledge Format (hsm export --format okf) plus Mem0/Zep: we're OKF, but signed and verifiable.
2. Verification over trust. Integrity is a number (RotBench, 0–100), computed by mechanical checks — no LLM judging its own homework. See benchmarks/ROTBENCH.md.
3. Auditable extraction. The optional distilled layer (docs/DISTILL_SPEC.md) extracts entity facts with verbatim quotes, checked in code — a claim either cites a real source or it's dropped. Contradictions append a changelog line (update current_crm: "Salesforce" -> "HubSpot" (source: chat-042.md)) — never a silent overwrite. Extraction you can audit is extraction you can trust.

The two camps (where this sits)

Honest numbers (LongMemEval)

Measured on the full 500-question _s set (48-session haystacks with distractors), scored with the official per-type judge methodology, reader glm-5.2, independent judge deepseek-v4-pro. Reproduce: benchmarks/README.md. Full run history including the failures: benchmarks/RESULTS.md.

Read the RotBench row carefully: it scores the vault this harness builds, so it is a statement about this run, not a promise that your vault will score 99. Pointed at a real working vault it usually will not, and that is the tool doing its job. See benchmarks/ROTBENCH.md for what the score does and does not measure.

What we will and won't claim: recall is elite and reader-independent; QA is honest and mid — published systems self-report higher on their own harnesses (Mem0 94.4%, Zep 63.8% independent); we publish the harness, the judge, and every failed experiment instead. No number here is from a harness you can't run yourself.

Design

• Cross-platform. Pure Python, stdlib-only core. CI: ubuntu / macos / windows.
• Degrades explicitly. qmd 2.1+ is optional. MCP failure falls back to the dedicated CLI; qmd failure falls back to a read-only scan. Machine-readable output names the engine and reason instead of silently pretending the fast path worked.
• Local by default. The HTTP API binds loopback with bearer auth + DNS-rebind protection; the MCP server is stdio (client-spawned). Nothing phones home.
• Temporal. Changelog lines make history queryable: hsm history note --as-of DATE.

Status

v0.2, building in public. Roadmap: ROADMAP.md. Break our benchmark: benchmarks/ROTBENCH.md — adversarial fixtures get merged.

MIT © Kinetic Labs Inc. · a FuckBigTech / HOMESTEAD project.