kioku-mesh 0.6.0

Cross-agent distributed memory over a mesh transport (currently Zenoh)

Shared memory for AI coding agents, across tools and machines.

kioku (記憶) means memory.

kioku-mesh gives coding agents a shared memory store. Claude Code, Codex CLI, Gemini CLI, and other MCP clients can save and search the same observations from one machine or from several machines on a trusted LAN/VPN mesh.

The default setup is local and needs no daemon. Mesh mode is available when you want the same memory pool replicated between hosts.

Why kioku-mesh

Coding-agent context gets fragmented across machines: which laptop did that work, what did the agent on the other host decide, and why does a secondary agent have to re-read everything from scratch just to give a quick second opinion? kioku-mesh keeps that memory in one shared pool so any agent, on any of your machines, can recall it.

Unlike long-term memory tools that store everything in one place, the shared pool is a peer-to-peer mesh you run yourself across your own machines (LAN / VPN / Tailscale) — no SaaS, no central account. The same memory is readable by Claude Code, Codex CLI, Gemini CLI, and any other MCP client.

Quickstart

uv tool install kioku-mesh
# or: pip install kioku-mesh

kioku-mesh init --mode local
kioku-mesh save "Chose Postgres over SQLite for analytics"
kioku-mesh search "Postgres"

Install the MCP server for your agent:

kioku-mesh mcp install --client claude-code
kioku-mesh mcp install --client codex-cli

The package installs two commands:

• kioku-mesh: the CLI.
• kioku-mesh-mcp: the stdio MCP server launched by your agent.

Modes

Mode	Use it when	Persistence	Extra service
local	You want memory on one machine	SQLite	none
hub	This machine is the always-on mesh hub	RocksDB	zenohd
spoke	This machine connects to a hub	RocksDB	zenohd

local is the default and the easiest starting point. Re-run kioku-mesh init --mode <mode> --force when you want to switch. For a short-lived Zenoh smoke test without provisioning anything, use kioku-mesh mesh start.

In mesh mode the Zenoh/RocksDB store is the source of truth, and each host's SQLite is a fast local read cache rebuilt from it — not a separate copy you have to reconcile. local mode is a standalone, SQLite-only setup for a single machine, so its saves live only in that local store and are not replicated to a mesh.

CLI

kioku-mesh save "Decided to keep billing events append-only" \
  --memory-type decision \
  --importance 4 \
  --subject billing

kioku-mesh search "billing events"
kioku-mesh get-memory <observation_id>
kioku-mesh delete <observation_id>
kioku-mesh gc --retention-days 30
kioku-mesh doctor

Useful environment variables:

Variable	Purpose
MESH_MEM_AGENT_FAMILY	Agent family, such as claude or codex
MESH_MEM_CLIENT_ID	Client name, such as claude-code
MESH_MEM_SESSION_ID	Optional stable session id
MESH_MEM_STATE_DIR	State directory; defaults under the user data dir
MESH_MEM_BACKEND	Override the backend selected by ~/.config/kioku-mesh/config.yaml; set local or zenoh
MESH_MEM_FORCE_REBUILD=1	Rebuild the local index at CLI startup
MESH_MEM_DISABLE_INDEX=1	Use the legacy Zenoh scan path instead of SQLite index
MESH_MEM_USER_ID	Your user slug for --visibility user (same value on all your machines)
MESH_MEM_TEAM_ID	Team slug for --visibility team
MESH_MEM_DEFAULT_VISIBILITY	Default scope for new saves: user, team, mesh (unset = legacy layout)

Visibility scopes (experimental)

save and the MCP save_observation tool accept --visibility / visibility to choose how far a memory replicates:

• user — only this user's machines (requires user_id, set via MESH_MEM_USER_ID or user_id: in ~/.config/kioku-mesh/config.yaml; use the same value on every machine you own)
• team — peers that host the team's storage (requires team_id)
• mesh — every peer on the mesh
• unset — follows default_visibility from config; with no config the legacy layout is used and nothing changes

The user/team ids are resolved from server-side configuration, never from tool arguments. See ADR-0019 for the design.

A per-project default can be set with a .kioku-mesh.yaml in the repository (found by searching upward from the current directory; env vars still win, the global config is the fallback):

# .kioku-mesh.yaml — project-local defaults
default_visibility: team
team_id: kioku-mesh

user_id is intentionally not read from this file (a committed file must never set a personal namespace), and every save response echoes the effective scope, e.g. saved: <id> (visibility=team/kioku-mesh).

MCP Clients

kioku-mesh mcp install handles the common setups:

kioku-mesh mcp install --client claude-code
kioku-mesh mcp install --client codex-cli

For Claude Desktop, Gemini CLI, ChatGPT Desktop, manual JSON/TOML examples, SessionStart hooks, and multi-agent identity recipes, see docs/mcp-clients.md and docs/multi-agent.md.

Multi-Host Mesh

Each host serves its agents from a fast local SQLite read index, backed by a Zenoh router + RocksDB store, and hosts replicate to each other over the mesh:

flowchart LR
  subgraph HostA["🖥️ Host A"]
    direction TB
    A1["Claude Code"]
    A2["Codex CLI"]
    AS[("SQLite index<br/>local read path")]
    AZ["zenohd<br/>Zenoh router + RocksDB<br/>(source of truth)"]
    A1 & A2 -->|"save / search"| AZ
    AZ -->|"subscriber · rebuild"| AS
    A1 & A2 -.->|"fast reads"| AS
  end
  subgraph HostB["🖥️ Host B"]
    direction TB
    B1["Codex CLI"]
    B2["Gemini CLI"]
    BS[("SQLite index<br/>local read path")]
    BZ["zenohd<br/>Zenoh router + RocksDB<br/>(source of truth)"]
    B1 & B2 -->|"save / search"| BZ
    BZ -->|"subscriber · rebuild"| BS
    B1 & B2 -.->|"fast reads"| BS
  end
  AZ <==>|"Zenoh mesh replication<br/>LAN / VPN / Tailscale · TCP 7447"| BZ

The recommended topology is one hub and any number of spokes. The hub listens on addresses reachable from the spokes; every spoke dials only the hub.

flowchart TB
  HUB["⭐ Hub<br/>always-on peer<br/>listens on LAN / Tailscale / VPN"]
  S1["Spoke · laptop"]
  S2["Spoke · desktop"]
  S3["Spoke · CI / server"]
  S1 -->|"dials hub (TCP 7447)"| HUB
  S2 -->|"dials hub"| HUB
  S3 -->|"dials hub"| HUB
  S1 -.->|"router transit<br/>(no direct link)"| S3
  S2 -.->|"router transit"| S3

# hub
kioku-mesh init --mode hub \
  --listen 127.0.0.1 \
  --listen 192.168.3.10

# spoke
kioku-mesh init --mode spoke \
  --listen 127.0.0.1 \
  --connect 192.168.3.10

Mesh mode requires zenohd and zenoh-backend-rocksdb on PATH. The current target is Zenoh 1.9.0.

zenohd -c ~/.config/kioku-mesh/zenohd.json5

For login auto-start on Linux systemd hosts, add --install-systemd to kioku-mesh init --mode <hub|spoke> ...; the wrapper writes a user unit pointing at the generated config.

kioku-mesh is designed to run inside a closed, trusted network. Keep port 7447/tcp reachable only between trusted peers. Do not expose it to the internet or an untrusted LAN. By default kioku-mesh relies on network admission (Tailscale, WireGuard, firewall rules, or a trusted LAN) rather than transport-level authentication.

When network admission alone is not enough, enable mutual TLS: every peer presents a certificate signed by your own private CA, and zenohd refuses any unverified link. Each peer's private key is generated locally and never leaves the host — only CSRs and signed certs (all non-secret) are exchanged.

kioku-mesh tls init-ca                            # once, on the CA host
kioku-mesh tls enroll <ca-host> --san <this-ip>   # on each peer (needs SSH to the CA host)
kioku-mesh init --mode <hub|spoke> --tls --listen ... --force

No SSH? The copy-paste flow works over any channel: tls request prints a CSR block, paste it into tls sign on the CA host, paste the bundle it returns into tls install. The peer key never leaves the host; only non-secret blocks move.

See docs/mtls.md for the full walkthrough, the trust model, and certificate rotation.

For a full walkthrough with firewall notes, five-peer examples, add/remove procedures, and smoke tests, see config/peers/example_5peer.md.

Development

pip install -e '.[dev,test]'
pytest tests/ -q

Run focused MCP checks with:

pytest tests/test_mcp_server.py tests/test_mcp_cli.py -v

Notes

• Python 3.10+ is required.
• Linux is the primary development and deployment target.
• Windows users should prefer WSL2. Native setup notes are in docs/windows-setup.md.
• macOS support is not verified yet.
• delete writes a tombstone. gc performs physical cleanup.
• 0.x releases are experimental; breaking changes can happen in minor versions.

More detail lives in docs/Spec.md, CHANGELOG.md, and the design records under docs/adr/.

Acknowledgments

kioku-mesh was influenced by engram and claude-mem. No code is copied from either project.