ClawMem

Name: ClawMem
Author: Yoloshi Nomotomoro?

Local, SQLite-backed memory and hybrid search over your markdown for coding agents

By Yoloshi Nomotomoro? (individual) JavaScript Local · stdio Knowledge and memory

GitHub npm Website

48 popularity score ★ 211 ↓ 190/wk

Connect your client

Local (npm)

Add in one click

Add to Cursor Add to VS Code Add to LM Studio

Opens the app, which asks before it adds the server.

Or add it by command or config

FirstTerminal

npm install -g clawmem
clawmem bootstrap <NOTES_DIR> --name <NAME>

Claude CodeTerminal

claude mcp add clawmem -- clawmem mcp

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "clawmem": {
      "command": "clawmem",
      "args": [
        "mcp"
      ]
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "clawmem": {
      "command": "clawmem",
      "args": [
        "mcp"
      ]
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "clawmem": {
      "type": "stdio",
      "command": "clawmem",
      "args": [
        "mcp"
      ]
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.clawmem]
command = "clawmem"
args = ["mcp"]

Gemini CLITerminal

gemini mcp add --scope user clawmem clawmem -- mcp

Local (from source)

Add in one click

Add to Cursor Add to VS Code Add to LM Studio

Opens the app, which asks before it adds the server. Then replace each <NAME> in its settings with your value.

Or add it by command or config

FirstTerminal

git clone https://github.com/yoloshii/clawmem.git <ABSOLUTE_PATH_TO_CLONE>
cd <ABSOLUTE_PATH_TO_CLONE> && bun install
<ABSOLUTE_PATH_TO_CLONE>/bin/clawmem bootstrap <NOTES_DIR> --name <NAME>

Claude CodeTerminal

claude mcp add clawmem -- '<ABSOLUTE_PATH_TO_CLONE>/bin/clawmem' mcp

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "clawmem": {
      "command": "<ABSOLUTE_PATH_TO_CLONE>/bin/clawmem",
      "args": [
        "mcp"
      ]
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "clawmem": {
      "command": "<ABSOLUTE_PATH_TO_CLONE>/bin/clawmem",
      "args": [
        "mcp"
      ]
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "clawmem": {
      "type": "stdio",
      "command": "<ABSOLUTE_PATH_TO_CLONE>/bin/clawmem",
      "args": [
        "mcp"
      ]
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.clawmem]
command = "<ABSOLUTE_PATH_TO_CLONE>/bin/clawmem"
args = ["mcp"]

Gemini CLITerminal

gemini mcp add --scope user clawmem '<ABSOLUTE_PATH_TO_CLONE>/bin/clawmem' -- mcp

Settings 10

Set together: embedding
`CLAWMEM_VAULTS`	Optional: JSON map of named vaults to SQLite paths. Without it there is one vault at ~/.cache/clawmem/index.sqlite. For list_vaults, vault_sync and the vault parameter.
`CLAWMEM_NO_LOCAL_MODELS`	Optional: Set to true to fail instead of running models in-process with node-llama-cpp. Default `false`.
`CLAWMEM_HEAVY_LANE`	Optional: Set to true for the quiet-window heavy maintenance worker. Meant for the clawmem watch service. Default `false`.
`CLAWMEM_EMBED_URL`	Optional: Embedding server address. Left unset, embedding runs in-process.
`CLAWMEM_EMBED_API_KEY` secret	Optional: Key for a cloud embedding provider. Setting it turns on cloud mode.
Set together: embedding wait
`CLAWMEM_EMBED_WAIT_SEC`	Optional: Seconds clawmem embed waits for a self-hosted embedding server when no cloud key is set and local models are off. Default `120`. For clawmem embed.
`CLAWMEM_EMBED_WAIT_INTERVAL_MS`	Optional: Milliseconds between probes during that wait. Default `5000`. For clawmem embed.
`CLAWMEM_EMBED_PROBE_TIMEOUT_MS`	Optional: Milliseconds one probe during that wait may take. Default `15000`. For clawmem embed.
Set together: consolidation
`CLAWMEM_ENABLE_CONSOLIDATION`	Optional: Set to true to run the consolidation worker inside the server. Default `false`.
`CLAWMEM_CONSOLIDATION_INTERVAL`	Optional: Consolidation worker interval in milliseconds. Default `300000`.

What it does

ClawMem keeps a long-term memory for coding agents on your own machine. It indexes collections of markdown (project docs, notes, research, Obsidian vaults) into a SQLite vault, stores decisions and handoffs taken from past sessions, and lets the agent search all of it with keyword (BM25), vector and combined reranked queries, plus graph lookups for causes, timelines and entity facts.

The MCP server is one of four ways in. The same vault is also fed by Claude Code hooks, a native OpenClaw memory plugin and a Hermes Agent memory provider, so what one runtime records is visible to the others. Its makers say the hooks do most of the work in Claude Code and the MCP tools are for targeted look-ups when the hooks miss something.

The stdio server is started with `clawmem mcp`. At v0.43.5 the server code (src/mcp.ts) registers 33 tools, including a `__IMPORTANT` guide tool that tells the agent which search to use. It also serves any indexed document as a `clawmem://{+path}` resource, which the README does not mention. Searches return 10 results by default (`limit`), `multi_get` caps output at 10,240 bytes by default (`maxBytes`), and `lifecycle_sweep` only previews unless `dry_run` is set to false. Embedding, query expansion and reranking run on local models through llama-server or in-process node-llama-cpp. Cloud embedding providers are an opt-in alternative.

What it needs

Runtime	Bun 1.0+
Runs on	Linux, macOS
Also	Homebrew SQLite on macOS (brew install sqlite) Optional GPU: about 4 GB VRAM for the default models, 16 GB for the larger stack bd CLI v0.58.0 or newer, only for beads_sync

On Linux, install Bun with the bun.sh script rather than snap, which breaks the hooks. Windows users run it inside WSL2. Before the server is useful, create a vault with content using `clawmem bootstrap <dir> --name <name>`. Local models download on first use.

Limits

Limit Without a GPU or llama-server, models run in-process and can fall back to CPU silently. Set CLAWMEM_NO_LOCAL_MODELS=true to fail instead.
Limit Contradiction detection only works when a judge model is configured with the CLAWMEM_JUDGE_* settings.
Not covered vault_sync refuses paths under /etc, /root, /var, /proc, /sys and /dev, and any path containing .ssh, .gnupg, .env, credentials, secrets, .aws or .kube.
Limit memory_forget deactivates the best match as soon as it is called (confirm defaults to true), and lifecycle_restore does not bring forgotten documents back.
Limit The makers warn that most zerank-2 GGUF files lack the score head and do nothing as rerankers.
Limit CLAWMEM_HEAVY_LANE is meant for the clawmem watch service. The MCP server warns that its quiet-window worker may never run there.
Limit From v0.43.5, clawmem embed against a self-hosted embedding server with local models off waits 120 s by default, then skips embedding and exits 1.
Platform Native Windows is not recommended because the launcher and hooks are bash scripts. Use WSL2.
Limit The README's MCP tools heading says 31, but the v0.43.5 server registers 33 tools.

Without a GPU or llama-server, models run in-process and can fall back to CPU silently; set `CLAWMEM_NO_LOCAL_MODELS=true` to fail instead. Contradiction detection only works when a judge model is configured (`CLAWMEM_JUDGE_*`). `vault_sync` refuses directories under /etc, /root, /var, /proc, /sys, /dev and paths containing .ssh, .env, credentials, secrets, .aws or .kube. The makers warn that most zerank-2 GGUF files lack the score head and do nothing as rerankers. The heavy maintenance lane (`CLAWMEM_HEAVY_LANE`) is meant for the `clawmem watch` service, and the MCP server warns that it may never run there.

Who it suits

Suits a developer who wants memory for Claude Code, OpenClaw or Hermes that stays on their own machine, works over their own markdown, and needs no API key. It is most useful with the hooks installed as well as the MCP server; with any other MCP client you get the 33 tools but no automatic capture or context injection. It asks more setup than a plain notes server: Bun, an indexed vault, local models and ideally a GPU. Pick something else if you are on native Windows, want a hosted memory shared across machines without running services, or need a store that agents only read: several tools write to the vault.

Tools 33

23 only read, 9 change things, 1 can delete or overwrite. A ? marks an optional parameter.

__IMPORTANT: Guide telling the agent which search tool fits which question
memory_retrieve: Classifies a query and sends it to the best-suited searchquery: stringmode?: enumlimit?: numbercompact?: booleanincludeInternal?: booleanvault?: string
search: Keyword (BM25) search of the vaultquery: stringlimit?: numberminScore?: numbercollection?: stringcompact?: booleanincludeInternal?: booleanvault?: string
vsearch: Vector search for loosely worded or conceptual queriesquery: stringlimit?: numberminScore?: numbercollection?: stringcompact?: booleanincludeInternal?: booleanvault?: string
query: Combined keyword and vector search with rerankingquery: stringlimit?: numberminScore?: numbercollection?: stringcompact?: booleandiverse?: booleanintent?: stringcandidateLimit?: numberincludeInternal?: booleanvault?: string
memory_forget deletes: Deactivates the closest matching memory at once unless confirm is falsequery: stringconfirm?: booleanvault?: string
profile: Returns the stored user profile, rebuilding it on requestrebuild?: boolean
get: Fetches one document by path or idfile: stringfromLine?: numbermaxLines?: numberlineNumbers?: booleanvault?: string
multi_get: Fetches several documents by glob or list of pathspattern: stringmaxLines?: numbermaxBytes?: numberlineNumbers?: booleanvault?: string
status: Reports index health and content typesvault?: string
find_similar: Finds documents near a given one in vector spacefile: stringlimit?: numberincludeInternal?: booleanvault?: string
reindex changes: Starts a re-scan of the vault's collectionsvault?: string
index_stats: Detailed index statistics: types, staleness, access counts, sessionsvault?: string
memory_stats: Per-collection lifecycle and ranking statisticscollection?: stringvault?: string
memory_rank: Explains how each result's ranking score was made upquery: stringlimit?: numbercollection?: stringweightProfile?: enumincludeInternal?: booleanvault?: string
session_log: Returns history of past sessionslimit?: numbervault?: string
beads_sync changes: Imports issues from a Beads tracker into the vault through the bd CLIproject_path?: string
build_graphs changes: Builds temporal and semantic graphs over the documentsgraph_types?: arraysemantic_threshold?: numbervault?: string
intent_search: Answers why/when/who questions by walking causal and semantic linksquery: stringlimit?: numberforce_intent?: enumenable_graph_traversal?: booleanvault?: string
query_plan: Splits a multi-part question into separate typed searchesquery: stringlimit?: numbercompact?: booleanincludeInternal?: booleanvault?: string
find_causal_links: Follows chains of decisions and their causesdocid: stringdirection?: enumdepth?: numbervault?: string
kg_query: Looks up subject-predicate-object facts about an entity, with time rangesentity: stringas_of?: stringdirection?: enumvault?: string
memory_evolution_status: Shows how a document's derived metadata changed over timedocid: stringlimit?: numbervault?: string
timeline: Lists documents created or changed before and after a given onedocid: stringbefore?: numberafter?: numbersame_collection?: booleanvault?: string
memory_pin changes: Pins a memory so it is kept and ranked higher, or unpins itquery: stringunpin?: booleanvault?: string
memory_snooze changes: Hides a memory from surfaced context until a datequery: stringuntil?: stringvault?: string
lifecycle_status: Counts active, archived, forgotten, pinned and snoozed documentsvault?: string
lifecycle_sweep changes: Archives stale documents by policy; previews by defaultdry_run?: booleanvault?: string
lifecycle_restore changes: Restores documents archived by lifecycle policiesquery?: stringcollection?: stringall?: booleanvault?: string
list_vaults: Lists configured named vaults and their paths
vault_sync changes: Indexes markdown from a directory into a named vaultvault: stringcontent_root: stringpattern?: stringcollection_name?: string
diary_write changes: Adds an entry to the agent diaryentry: stringtopic?: stringagent?: stringvault?: string
diary_read: Reads recent diary entries, optionally for one agentlast_n?: numberagent?: stringvault?: string

Resources 1

clawmem://{+path}: One indexed document as markdown with line numbers, addressed by collection and path

Popularity

48 popularity score, on our list (the bar is 36)

★ 211 ↓ 190/wk

GitHub stars	211	27.9 points
Stars in 30 days	+1	1.5 points
npm downloads, last week	190	18.2 points

Measured 2026-11-07. How the score works.

Where it is published

Code	yoloshii/clawmem ★ 211
npm	clawmem 0.43.5, 2026-10-20

What it can reach

Runs	On your machine
Reaches	Files on your machine, A database
Changes things	Yes: it can write, send, create or delete
Sign-in	None
Licence	MIT (npm)
Last code change	2026-10-20

Runs on your machine over stdio and opens no network port; the separate `clawmem serve` REST API listens on localhost:7438 and needs a token. Data sits in local SQLite files (default ~/.cache/clawmem/index.sqlite). Tools read indexed markdown and can write: `vault_sync` indexes any directory outside a short deny-list, `memory_forget`, `memory_pin`, `memory_snooze`, `lifecycle_sweep` and `diary_write` change the vault, and `beads_sync` runs the bd CLI. Models run locally through llama-server (localhost:8088-8090 via the launcher) or in-process, downloaded on first use; text leaves the machine only if you configure a cloud embedding provider. The makers say surfaced content is screened for prompt injection.

Over time

GitHub stars 211 +1

213210

9 Oct 20267 Nov 2026

npm downloads a week 190 −581 (−75 %)

2.5k24

23 Mar 20267 Nov 2026

Before the site measured it, read from npm's daily download counts.

Popularity score 48 −6 (−11 %)

5444

9 Oct 20267 Nov 2026

What is rising across the directory

Sources

Page checked 2026-10-30.

Badge for your README

Shows this server's rank in its category, and stars for a server that runs on your machine. It updates every hour.

MarkdownREADME.md

[![ClawMem on EveryMCPServer](https://everymcpserver.com/badge/clawmem.svg)](https://everymcpserver.com/servers/clawmem)

Connect your client

Add in one click

Or add it by command or config

Add in one click

Or add it by command or config

Settings 10

What it does

What it needs

Limits

Who it suits

Tools 33

Resources 1

Similar servers

Clawdi

Engram

projectmem

agentmemory

marm-memory

SuperLocalMemory