Facthouse

Facthouse is a local memory engine for AI tools. Most “memory” products index chat logs. Facthouse takes agent activity - messages, tool use, and other MCP traffic - and applies neuroscience-inspired consolidation so it moves through Data (what happened in the session) → Information (extracted facts) → Knowledge (integrated beliefs on an entity graph). During this process, Facthouse links entities, drops duplicates, reconciles conflicts, and supersedes what is out of date. Vector embeddings add optional semantic search on top of that graph. The store is a SQLite file on your disk.

Quick Start

Needs Node 22.5 or 24+.

npx -y @facthouse/mcp
facthouse init

facthouse init --web is the same setup as a browser form — it prints a 127.0.0.1 URL and does not open a browser.

Press Enter to accept each default (copy = Claude Code or Cursor session logs on disk; type record if the assistant should save facts). If you picked copy, init asks whether to copy existing logs, then whether to extract and integrate. Init prints an MCP snippet as soon as the store is written — add it to the client's MCP config while copy/extract run. Restart the client when init finishes.

In the client, state something durable in ordinary conversation — there is no remember command.

Ask it back in the next session, or facthouse search. That is the store.

Copy from Claude Code or Cursor logs, or record from any MCP client: How conversations get in. Replay: facthouse.dev/demo.html. CLI: below.

What you get

• Local SQLite. Optional Postgres. Isolation is the directory, not a column.
• Entity graph. People, organisations, projects, places, products — extracted, typed and linked.
• Hybrid search. BM25 + structured domain + entity-graph paths, merged via Reciprocal Rank Fusion. An embedding provider adds meaning as a fourth list; off by default.
• In-session memory. get_session_context is the same briefing as memory://briefing. Tools-only clients should call it at session start.
• Immutable history. Facts are never deleted, only superseded.

How it works

One SQLite database. Three tables in it, not three databases: Data (what happened in the session) → Information (extracted facts) → Knowledge (integrated beliefs).

• D (session_events) — what was said (copied transcripts, or what the assistant records)
• I (session_facts) — what was just extracted, or capture_fact
• K (facts) — integrated knowledge

FTS5 (words) and optional embeddings (meaning) are indexes of K. They are not a second store. Semantic search is off unless you turn it on: search "shellfish" finds a shellfish fact, search "food" does not, until you choose an embedding model — a model is an opinion about what “similar” means.

Two speeds. Copy tails named transcripts into Data. Extract turns new transcript lines into self-contained facts (D→I). Integrate fits them into what the store already knows: domains, entities, duplicates, contradictions, the graph (I→K). consolidate is the umbrella: copy, extract, and integrate together. Extract is capped at 50 lines per run, so a first backfill is never spent on the lot; when extract runs, it takes the oldest 50 lines. Consolidation does not invent a sentence nobody said.

A hook cannot call MCP tools — those exist only on the assistant’s connection — and it must not wait for a model pass. So it does not invoke consolidate. It runs facthouse notify …, which tells the already-running server that a moment happened and returns at once. consolidate is the pipeline verb (MCP tool or CLI); the caller waits. notify compaction is that verb asked of the live server, asynchronously. notify threshold is a different policy (extract only, if due).

Automatic

Callable

On a default copy store with no extra hooks, only the automatic table runs: Facthouse starts (all three), copy on each Facthouse tool or resource call if named sources grew, and integrate on a clean process exit. Closing a chat window may skip the exit row; the next start still consolidates. Due on threshold means at least 10 unexamined lines and two minutes since the last gated run. consolidate --all lifts the extract cap. Compaction is recommended PreCompact (notify compaction): same three steps as consolidate, on the running server, without waiting. We do not install the hook. Not a Stop hook. record wakes threshold extract on a record store — do not install record hooks on a copy store.

Storage needs Node. Intelligence needs a language model. By default that is the Claude Code CLI on your existing subscription. Without it, consolidation falls back to a built-in heuristic that does not extract facts from transcripts. capture_fact still stores facts, with no entities and no domain routing.

How conversations get in

Two ways. Pick one per store.

On a copy store, capture_fact is a correction for every MCP client, not only the one that writes JSONL. Grok has no transcript adapter — do not put Claude Code on copy and Grok on the same store expecting Grok to record.

facthouse init

Pick copy, set cwd. Init asks whether to copy existing logs, then whether to extract and integrate (Enter = all copied lines; a selection of 500 or more asks you to type the choice again). Decline extract to do that later with facthouse consolidate (--all takes remaining unexamined lines, not ones skipped as outside a 7-day or 30-day window). After that, the server copies new lines when it handles a Facthouse call. Extract and integrate follow the table in How it works.

Compact (recommended): facthouse notify compaction — we do not install the hook. Not a turn-end Stop hook.

MCP

Works with any MCP-compatible tool. Default store: ~/.facthouse. A different path is "env": { "FACTHOUSE_DATA": "/absolute/path" } on the MCP snippet. JSON accepts forward slashes on Windows.

Cursor consumes tools but not resources until a later adapter exists — search_knowledge and get_entity still work there; call get_session_context at session start.

Resources are context the client loads automatically — no tool call. Tools only help if the assistant remembers to reach for them; resources are simply present.

• memory://briefing — Everything worth knowing right now: profile, what was learned in the last consolidation, open threads, and recent knowledge. Markdown, kept to roughly a screenful.
• memory://profile — Core identity facts, most important first.

Both are read-only views over the same database the tools query. Clients that never load resources (Cursor, Windsurf, Grok) get the same briefing by calling get_session_context at the start of a conversation. No second profile schema.

Tools

Session

• log_event — Log conversation events (messages, artifacts).
• get_events — Retrieve events from current or previous session.
• get_session_context — Working briefing (the same markdown as memory://briefing) plus facts captured in this session. Call at the start of every conversation if the client does not load resources.

Reading

• get_entity — Everything known about any named subject — person, organisation, project, place, product — and how it connects. When several rows share the name under different types, facts from all of them come back. Hyphens, underscores, and stray punctuation count as the same letters only when that does not join two names already stored as separate rows. If there is no entity by that name, facts that mention the wording still come back rather than an empty miss.
• get_context — Everything relevant to a topic (search + entity traversal)
• search_knowledge — Hybrid search across integrated knowledge

Writing

• capture_fact — Store a fact. On a copy store this is a correction for something extraction missed; on a store with empty sources it is how facts get in. The description the assistant sees is generated from that same rule.
• consolidate — Integrate pending facts into long-term knowledge. Extracts entities, resolves duplicates, detects contradictions, builds the knowledge graph.
• Inference tools — Opt-in, off by default (inferences.enabled in config.json). A hypothesis cites existing fact ids and stays pending until confirmed. Those tools are not registered until you turn the gate on. Consolidate never invents a sentence nobody said.

Meta

• get_schemas — Available domains and structure
• get_stats — Fact count, entity count, domain distribution, extract backlog, intelligence spend

CLI

The MCP JSON starts the server via npx and does not need a global install. npm install -g puts facthouse on PATH for init, settings, stats, and inspect. The same CLI without PATH is npx -y -p "@facthouse/mcp" -- facthouse — pin the version; quote the package so PowerShell does not splat. -p and -- stop an older global binary winning. npx -y @facthouse/mcp with no -p / facthouse is the server; do not run it as a shell command for init, settings, or stats. The MCP paste starts the server. It does not put facthouse on PATH. To inspect the file from a terminal, see CLI below.

These CLI commands work in bash, zsh, and PowerShell. Quote @facthouse/mcp in PowerShell. Git Bash /c/... paths are not PowerShell; use C:/... and pass --data instead of cd or export. In Git Bash, quote a backslash path or write C:/... — unquoted \ is an escape. ~/ is expanded on every platform. WSL uses /mnt/c/.... FACTHOUSE_DATA on an MCP snippet applies only to that server process. A terminal facthouse command needs --data, FACTHOUSE_DATA in the environment that shell inherits, or a .facthouse store in this project. Hooks do not see mcp.json env.

npm install -g @facthouse/mcp@0.31.0
facthouse init --yes

If npm install -g fails because a command named mcp already exists, remove that leftover command and retry.

npx -y -p "@facthouse/mcp@0.31.0" -- facthouse init --yes
npx -y -p "@facthouse/mcp@0.31.0" -- facthouse settings --json
npx -y -p "@facthouse/mcp@0.31.0" -- facthouse stats
npx -y -p "@facthouse/mcp@0.31.0" -- facthouse inspect

facthouse init [dir]

The walk-through is how a human first-run writes config.json. Skip it and the server still creates the directory on first MCP boot.

On a terminal, init asks data directory, copy transcripts vs assistant records (default copy), semantic search, and More settings. --yes never prompts and leaves sources empty. --web prints a 127.0.0.1 URL and does not open a browser; --yes refuses --web. On a terminal, --force still asks those questions, then replaces the whole file; --yes --force is the silent reset. --force does not merge with the previous file.

facthouse init --yes

facthouse init --yes ~/my-memory

facthouse init --yes --force

The generated config.json is where you change consolidation behaviour — most notably intelligence.provider (cli by default; heuristic for a zero-dependency regex fallback, or FACTHOUSE_PROVIDER=heuristic at runtime). Init does not ask that field.

facthouse settings

Change extra knobs on an existing config.json (CLI model, timeout, optional local extract). Does not reset the rest of the file. Refuses if there is no config.json (this command does not create a store). --json / not a terminal prints the current knobs and does not write. --web is the same knobs on a local page (print URL, no auto-open).

facthouse settings

facthouse settings --data ~/my-memory

facthouse record

Inserts events directly into the database (no running server needed). Supported for demos and for stores that have no named source. Not the Claude Code or Cursor default — that is sources plus facthouse consolidate.

# From a hook (reads JSON payload from stdin):
echo '{"hook_event_name":"UserPromptSubmit","prompt":"hello"}' | facthouse record --role user

# With explicit content:
facthouse record --role user --event-type message --content "hello world"

# Options:
#   --role          user | assistant | system | tool (default: user)
#   --event-type    message | tool_call | tool_result | artifact (default: message)
#   --content-type  text | json | image | audio | binary (default: text)
#   --content       Event content (or pipe via stdin)
#   --speaker       Named participant when the transcript has one
#   --session-id    Target session (default: most recent)
#   --data          Data directory (default: FACTHOUSE_DATA, a .facthouse store in this project, or ~/.facthouse)

facthouse consolidate

Copy new lines from config.sources, extract candidate facts from them, and integrate the pending facts into knowledge. The one command that spends model calls:

facthouse consolidate
facthouse consolidate --copy            # copy only; spends nothing
facthouse consolidate --integrate       # pending facts to knowledge; no extract pass
facthouse consolidate --all             # extract the whole backlog now
facthouse consolidate --limit 200       # extract the oldest 200

# Steps — named steps run, in order; none named means all three:
#   -c, --copy       copy new transcript lines into the store
#   -e, --extract    turn new lines into candidate facts (the model call)
#   -i, --integrate  classify, link, dedupe, supersede, embed
# Extract is capped at 50 lines per run so a first backfill is never spent on
# the lot; the run says how many remain. --all lifts the cap, --limit N sets it.
#   --json           print the result object instead of the summary
#   --data           Data directory (default: FACTHOUSE_DATA, a .facthouse store in this project, or ~/.facthouse)

Honours the configured provider (by default claude -p). Empty sources makes the copy step a no-op. Set cwd on the source unless you intend to copy every project group. Do not also run record hooks on a store with named sources.

facthouse notify <moment>

Tell the running MCP server that a moment happened. The server decides what to run and does it in the background, so a hook returns at once:

facthouse notify compaction   # client about to compact: copy new JSONL, extract, integrate (does not wait)
facthouse notify threshold    # events arrived: extract if the threshold is due

# Options:
#   --data     Data directory (default: FACTHOUSE_DATA, a .facthouse store in this project, or ~/.facthouse)

No server listening is not an error: the command says so and exits 0, and the next session start covers it. This is what the PreCompact hook calls.

facthouse search <query>

facthouse search "coffee"
facthouse search "coffee" --domain preferences
facthouse search "coffee" --json

# Options:
#   --domain   Prioritise a domain. Biases ranking; does not filter
#   --limit    Maximum results (default: 20)
#   --json     Emit the raw search payload
#   --data     Data directory (default: FACTHOUSE_DATA, a .facthouse store in this project, or ~/.facthouse)

--domain biases ranking rather than filtering. A hard filter would hide a fact filed under a near-synonym.

facthouse stats

facthouse stats
facthouse stats --json

Facts are immutable — superseded facts are kept — so the current count and the total legitimately differ once anything has been superseded. --json includes the answering binary's package version. Intelligence spend is calls, tokens, and elapsed time for extract / classify / entities / reconcile / supersede / summarise, with provider and model per stage. Embeddings are not that number.

facthouse inspect

Sample D, I, K, entities, and the graph. Writes a local HTML file under the data directory (not the cwd). Prints the path. Does not open a browser. The file is a memory export — treat it like stats --json. The same page also shows intelligence spend (Graph / Spend).

facthouse inspect
facthouse inspect --graph
facthouse inspect --layer k
facthouse inspect --json
facthouse inspect --entity Helios --limit 20 --output ~/inspect.html

--layer health|d|i|k|entities|graph|all prints terminal tables (newest-first, capped). --graph (the default when no --layer / --json) writes inspect.html. --limit is 10 for tables and 50 for the canvas. --all draws every node — a hairball, explicit. Search and type filter in the page can still reach a node that was outside the cap.

Advanced

Another store

The store is this directory. Clients share it by using the same path. A second store is a second directory, not a second install. The default MCP server name is facthouse. Splitting is not a filter on which client wrote the row. Work and personal is one reason to split, not a required setup.

A non-default data directory prints a distinct MCP server name so two stores can share one mcp.json. Init against each extra directory prints that snippet. Example:

{
  "mcpServers": {
    "facthouse-personal": {
      "command": "npx",
      "args": ["-y", "@facthouse/mcp@0.31.0"],
      "env": { "FACTHOUSE_DATA": "C:\\Users\\alex\\.facthouse-personal" }
    },
    "facthouse-work": {
      "command": "npx",
      "args": ["-y", "@facthouse/mcp@0.31.0"],
      "env": { "FACTHOUSE_DATA": "C:\\Users\\alex\\.facthouse-work" }
    }
  }
}

Point each store's sources.cwd (or hook --data) at that store only. Two directories do not isolate anything if both copy the same home.

Postgres (optional)

SQLite is the default and needs no extra software. To use Postgres instead, set storage.provider to "postgres" in that store's config.json, or FACTHOUSE_STORAGE=postgres on the MCP entry, and set FACTHOUSE_POSTGRES_URL to a postgres:// (or postgresql://) URL. The password belongs in the environment, not in config.json. If the URL is missing or the server cannot be reached, Facthouse stops; it does not create a SQLite file.

The data directory is still the memory: config.json and the scheduler socket live there. Tables live at the URL. Two memories need two directories and two databases.

Init does not ask which engine to use. facthouse init --yes still writes sqlite.

Example — placeholders only; do not put a real password in a committed file:

{
  "mcpServers": {
    "facthouse": {
      "command": "npx",
      "args": ["-y", "@facthouse/mcp@0.31.0"],
      "env": {
        "FACTHOUSE_DATA": "C:\\Users\\alex\\.facthouse-work",
        "FACTHOUSE_STORAGE": "postgres",
        "FACTHOUSE_POSTGRES_URL": "postgres://USER:PASSWORD@localhost:5432/facthouse"
      }
    }
  }
}

Copy versus record

Choose one mechanism per store.

Recommended — copy. Name a claude-code or cursor source (set cwd) and run facthouse consolidate from the CLI first. The MCP server also copies at session start and when it handles a call. Grok and Codex are later adapters. Unknown kind values are rejected.

{
  "sources": [
    {
      "kind": "claude-code",
      "home": "~/.claude",
      "cwd": "C:\\dev\\app"
    }
  ]
}

home is the client config dir (~/.claude or ~/.cursor — path examples, not extra discovery). Cursor is "kind": "cursor" and home/projects/*/agent-transcripts/**/*.jsonl only — not Composer SQLite. Cursor encodes C:\\dev\\app as c-dev-app (Claude Code uses C--dev-app). A first backfill of more than 50 lines takes several runs, or one facthouse consolidate --all.

Alternative — record, no sources. Leave sources empty. Pipe a client hook payload into facthouse record if you have one. MCP log_event / capture_fact keep working.

Do not install record hooks on this store — both write the same rows. Facthouse does not detect or rewrite existing hook configs.

MCP-only record mode

To skip the wizard (record only — no transcript copy), paste this. The server creates ~/.facthouse on first boot; you are not asked those questions.

{
  "mcpServers": {
    "facthouse": {
      "command": "npx",
      "args": ["-y", "@facthouse/mcp@0.31.0"]
    }
  }
}

Hooks

PreCompact notify compaction is the useful one: when the client is about to compact, a short-lived hook tells the running server, and the hook returns at once. The server copies new JSONL lines, then extracts and integrates — compaction does not delete the transcript, and the hook does not photocopy the live window. Init prints this JSON with --data filled in. Paste into Claude Code .claude/settings.json (user or project). We do not install it. Do not install a Stop hook. Do not install record hooks on this store — both write the same rows.

mcp.json env is not visible to hooks. Pass the same --data (or set FACTHOUSE_DATA in the environment the client itself inherits). The command must invoke the CLI (facthouse), never the server binary. npx -y @facthouse/mcp with no -p / facthouse starts the MCP server and hangs a hook. Pin the package version, quote it if the hook runs PowerShell, and put -- before facthouse so a globally installed older binary on PATH cannot win.

{
  "hooks": {
    "PreCompact": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "npx -y -p @facthouse/mcp@0.31.0 -- facthouse notify compaction --data /absolute/path/to/the-same-store"
          }
        ]
      }
    ]
  }
}

PreCompact notify compaction asks the running server to consolidate: copy the newest JSONL lines, extract, integrate. The hook returns at once; the server does the work. We do not install a turn-end Stop hook. On Windows the --data path is the same absolute directory you put in FACTHOUSE_DATA (for example C:\\Users\\alex\\AppData\\Local\\Temp\\facthouse-try).

Frequent incremental copying interleaves conversations on the global sequence: a long chat kept open is sliced between other chats. Extract progress is per conversation, so a timeout in one chat does not discard another. Shrinking extraction.batch_size means more extract calls (more chances of a timeout), not a store-wide hold-all. facthouse stats reports unextracted events against that extract watermark.

If the MCP server does not start, or lists no tools, check the package version the client actually spawned. A global facthouse on PATH can be years behind the pin in this README. Diagnose with facthouse stats --data <dir> (the CLI prints whether the scheduler is listening) and by inspecting serverInfo.version from initialize plus tools/list over stdio. 0.2.x answers initialize then throws on tools/list.

Embeddings, model, timeout, bitemporal

Set embedding.provider in config.json to "ollama" (local, no API key) or "voyage" (hosted), run facthouse consolidate, and search "food" starts returning the allergy. Facts are embedded when they are consolidated. Voyage applies a 3 requests/minute rate limit until a payment method is on the account.

Meaning-search is an exact scan of stored vectors when the set is small. When that set is large (default 32 MiB of the current model), an HNSW index of those vectors is used instead: in-process on SQLite, or a Postgres vector sidecar when the extension is enabled. Small stores stay exact. A missing engine keeps exact search and prints a warning; Facthouse does not install a native addon. embedding.ann is null (auto), false (never), or true (force when the engine allows). This does not turn embeddings on.

intelligence.cli.model and intelligence.cli.timeout_ms are extra knobs. First-run More settings (Y) can write them; later, facthouse settings. Init does not ask intelligence.provider; FACTHOUSE_PROVIDER=heuristic is the kill-switch. The heuristic fallback does not extract facts from transcripts.

Unnamed user-channel speech is attributed to the store's owner; a display name still does not create a person. Extra backing (assent, a tool observation, a different speaker restating) is recorded, not scored, unless the store sets interlocutor ranking weights in config.json. The engine ships none. Weight keys match the speaker string as stored, so two people with the same name share a key.

Set temporal.mode to bitemporal to record when the system retracted a belief, so search can answer what the store believed at an instant.

Intelligence spend

facthouse stats and get_stats report billed consolidation calls: tokens, elapsed time, and the provider plus model on each stage (extract, classify, entities, reconcile, supersede, summarise). A run that did not report tokens omits those fields rather than showing zero. Embeddings are a different API and are not this number.

Optional intelligence.token_budget caps billed extract per provider on rolling windows. Unset is unlimited. Over the cap, consolidate skips extract, holds the watermark, and does not fall back to the heuristic. Stats and inspect Spend show used and remaining on each cap, and when oldest usage in that window ages out (resets).

"intelligence": {
  "token_budget": {
    "cli": { "week": "10M" }
  }
}

hour, day, week, and month are rolling. Omit a scale to leave it unlimited. Remaining room is on facthouse stats, get_stats, and inspect Spend. Set the cap in this store's config.json — there is no budget command.

Optional local intelligence is a different switch from embeddings. Add intelligence.http on an OpenAI-compatible host. The protocol is POST /v1/chat/completions; only the port changes:

The model string is whatever that host lists. GET {base_url}/models prints the names. nomic-embed-text is embed-only and will not extract. If the host is up and serves exactly one chat model, Facthouse uses it for this run and tells you to pin intelligence.http.model. If several chat models are listed, set that field; extract will not guess.

Extract and summarise then use that host; reconcile and supersede stay on the CLI unless you list intelligence.stages. Each stage can set on-fail to cli, http, or none (see the JSON below). HTTP extract defaults to retrying on the CLI (counts against the CLI token budget). Contradiction defaults to none — no provider switch. none holds the extract watermark — it does not fall through to the heuristic. First-run More settings (Y, after the recommended path) can set the host, model, and extract on-fail. Later, facthouse settings merges those knobs into an existing file without resetting it. facthouse inspect Spend shows the same knobs and copies JSON; it does not save config.json.

The live script npm run test:http-intelligence has passed on qwen2.5vl:7b.

"intelligence": {
  "http": {
    "base_url": "http://localhost:11434/v1",
    "model": "qwen2.5vl:7b"
  },
  "stages": {
    "extract": { "provider": "http", "on_fail": "cli" },
    "summarise": { "provider": "http", "on_fail": "cli" },
    "reconcile": { "provider": "cli", "on_fail": "none" },
    "supersede": { "provider": "cli", "on_fail": "none" }
  }
}

CLI demo (no transcript source)

Throwaway store, not the capture path for a real Claude Code or Cursor home. These three lines are typed in.

export FACTHOUSE_DATA=/tmp/facthouse-demo
om() { npx -y -p "@facthouse/mcp@0.31.0" -- facthouse "$@"; }

om init --yes

om record --role user --content "I prefer dark mode in every editor, and I never want telemetry enabled."
om record --role user --content "I am allergic to shellfish, so avoid seafood restaurants when booking anything."
om record --role user --content "My colleague Robin at Acme is leading the Atlas migration project this quarter."

om consolidate
om search "Atlas"
om stats

$env:FACTHOUSE_DATA = Join-Path $env:TEMP "facthouse-demo"
function om { npx -y -p "@facthouse/mcp@0.31.0" -- facthouse @args }
om init --yes
om record --role user --content "I prefer dark mode in every editor, and I never want telemetry enabled."
om record --role user --content "I am allergic to shellfish, so avoid seafood restaurants when booking anything."
om record --role user --content "My colleague Robin at Acme is leading the Atlas migration project this quarter."
om consolidate
om search "Atlas"
om stats

allergies is not a domain Facthouse ships. The engine has no built-in vocabulary — it read the conversation and decided that fact needed a home. A domain biases ranking rather than filtering. Clean up: rm -rf /tmp/facthouse-demo (Git Bash / macOS / Linux) or Remove-Item -Recurse -Force $env:TEMP\facthouse-demo (PowerShell).

Integration

Facthouse's tool descriptions tell assistants when to search and when a correction is worth staging. They are not how Claude Code conversations enter the store — that is copy from a named source.

Without configuration

Claude Code or Cursor: name a sources entry (set cwd) and run facthouse consolidate from the CLI first. MCP session start also copies. capture_fact is there if the assistant needs to correct or add something copy-plus-extraction will not produce.

Clients with no copy adapter still rely on log_event / capture_fact until their adapter exists.

Hook points

On pre-compaction: facthouse notify compaction asks the server to consolidate. It is not a record hook.

Claude Code

Create .claude/rules/facthouse.md in your project (or ~/.claude/rules/facthouse.md globally):

# Facthouse

- Conversations are copied from the named Claude Code source (first backfill: `facthouse consolidate` on the CLI)
- Do not install record hooks on this store — both write the same rows.
- Identity context loads automatically from the `memory://profile` resource — no tool call needed
- Before answering questions this store might already know, call `search_knowledge`
- Call `capture_fact` only to correct or add something that is not in the transcript
- When the conversation is getting long, call `consolidate` (or rely on PreCompact `facthouse notify compaction`)
- At natural breakpoints (topic change, task completion), call `consolidate` to keep the knowledge graph current

To allow Facthouse tools without per-call approval prompts, add to the permissions.allow array in .claude/settings.json:

{
  "permissions": {
    "allow": [
      "mcp__facthouse__*"
    ]
  }
}

Cursor / Windsurf

On Cursor Directory, Add to Cursor is per component: click it on the MCP server and again on the rule. That boots a record store at ~/.facthouse — no facthouse init. The Hooks tab is Copy into ~/.cursor/hooks.json, not Add to Cursor. To copy Agent transcripts, run facthouse init, pick copy, kind cursor.

Add to .cursorrules (Cursor) or .windsurfrules (Windsurf) in your project root:

When the facthouse MCP server is available:
- At the start of every conversation, before answering, call get_session_context unless you already loaded the memory://briefing resource. That call returns the same working briefing the resource would have injected. Tools-only clients never fetch resources.
- Before answering questions this store might already know, call search_knowledge
- To find out everything known about a particular person, project, or thing, call get_entity
- Call capture_fact only to correct or add something copy or extraction missed

Cursor and Windsurf consume tools but not resources, so memory://profile will not load on its own there. Cursor conversations themselves are copied with kind: "cursor" (JSONL under ~/.cursor/projects/, not the SQLite composer store).

Claude Desktop / other MCP clients

No copy adapter yet. Tool descriptions handle search and optional capture_fact; conversations are not tailed until a later adapter exists.

Reclaiming space

Facthouse logs raw conversation and tool output to session_events. On a store wired into an agentic client this becomes almost all of the database. A store measured in daily use held 47,000 events and 493 MB against 21 integrated facts.

facthouse stats reports the raw layer alongside the knowledge, including how much is reclaimable. To reclaim it:

facthouse prune                    # report only — nothing is deleted
facthouse prune --apply --vacuum   # delete, then rebuild the file

Set retention.disk_budget in config.json to a size such as "2GB" to cap memory.db. Unset is unlimited; init does not write a cap. When a cap is set and the file is full, unreachable raw events are pruned automatically so new logs can reuse that space; if nothing unused remains, more raw events are refused. Facts are never deleted to meet the number. Compacting (--vacuum) is still a human step — it copies the whole file so the operating system sees the smaller size.

If most of that volume is tool output you judge to be noise, extraction.event_types and extraction.roles restrict what is examined, and extraction.min_content_length skips trivial events. Measure before you do. Volume and value are not the same axis.

The rule is reachability, not age. An event is removed only when all three hold:

1. Extraction has already read it. Anything ahead of the consolidation watermark is still input.
2. No fact's provenance cites it.
3. It has fallen outside its own session's most recent extraction.working_memory_size events — a spare so consolidation can still glance at recent raw notes. That window is evidence of the current topic, not a pronoun dictionary.

No fact, entity, embedding or search result is affected. Deleting rows does not shrink the file on its own — that is --vacuum. Without a cap, nothing prunes automatically.

Development

git clone https://github.com/gordonkjlee/facthouse
cd facthouse
npm install
npm run build
npm test

npm test always runs hermetic pipelines (fixture JSONL → copy → extract → search) with a recording extractor, and skips live evals that need a real model:

• Semantic recall needs Ollama with nomic-embed-text. Start it, then npm run test:semantic.
• The live first-fact eval needs the claude CLI. Run npm run test:first-fact.
• The live coding-store eval (warehouse-shaped Cursor transcripts) also needs the claude CLI. Run npm run test:coding-store.
• Local HTTP extract needs a chat model on an OpenAI-compatible host and FACTHOUSE_HTTP_MODEL (verified on qwen2.5vl:7b). Run npm run test:http-intelligence.

Each of those scripts fails rather than skips when its dependency is missing, so a green run means the claim was actually verified rather than quietly stepped over.

Contribute

Issues and pull requests are welcome. Open an issue first if the change is more than a typo.

• Questions: GitHub Discussions
• How to build and test: CONTRIBUTING.md

License

MIT