The candidate installer provisions a business memory service on its owner's Mac. It creates a dedicated database and restricted roles, stores database credentials in macOS Keychain and binds a coding project to the local service. Complete two-account installation, secure transfer and recovery acceptance remain release gates. Use fictional information during testing; this is not a production installation approval.
| Requirement | Purpose |
|---|---|
| Apple Silicon Mac | First native test target; Intel is unverified |
| Separate operator and coding Mac accounts | Keep publication and database administration outside unrestricted coding agents' reach |
| Python 3.11 or newer | Run the installer using the standard library |
| Node.js 24 | Run the packaged service; no npm install is required |
| PostgreSQL 18 tools | Create and maintain the dedicated database |
| Git or a downloaded release checkout | Obtain the software |
| Your coding framework and authenticated model tools | Run coding sessions with their existing models |
Use a trusted operator account to maintain the software. Run unrestricted coding agents only in the separate coding account. A separate terminal, environment variable, Keychain label or container under one UID does not provide this boundary. Setup refuses the same account, root execution and software that the coding account can alter. It does not create Mac accounts or grant administrator permissions.
The coding account must be standard, without administrator/root group membership, custom sudo rules or another way to assume the operator's identity. Administrator accounts remain trusted. Software and data must live on a filesystem that enforces ownership; volumes configured to ignore ownership are refused.
The operator must remain logged in, with the required Keychain accessible, while the coding account uses the service. This is a user LaunchAgent, not a system daemon. Fast user switching and locked-Keychain behaviour need acceptance on the target Mac. A permanently logged-out operator is not a supported unattended arrangement.
Keep the existing three-model YOLO workflow. Storage, access checks and lexical reads do not make inference calls; the coding and review workflow retains its model requirements.
If Homebrew is absent, follow its official installation guide. The formula route targets supported Apple Silicon systems. From the operator account:
brew install python@3.14 node@24 postgresql@18
export PATH="$(brew --prefix node@24)/bin:$(brew --prefix postgresql@18)/bin:$PATH"
python3.14 --version
node --version
psql --version
pg_dump --version
pg_restore --versionThe Python, Node and PostgreSQL formula pages document these install commands. Versioned Node/PostgreSQL formulas are keg-only, so the PATH step matters. Setup accepts explicit --node and --pg-bin locations, validates their resolved executables, and discovers a suitable Python without changing global PATH or shell files.
Do not run brew services start postgresql@18 for this kit. Its database has separate storage and lifecycle. Homebrew may create a default cluster; the installer does not reuse it or repoint an existing database.
Obtain the reviewed candidate checkout and enter its root. The public default branch remains the previous file-vault kit until this candidate passes acceptance. Do not import a personal vault as an installation step.
These examples use the fictional workspace demo-business and an already-created standard Mac account memory-coder. Replace both deliberately. Workspace names contain 2–40 lowercase letters, digits or hyphens, starting with a letter; 32-character hexadecimal identifiers are refused.
node runtime/team-memory-verify.mjs
./memory-kit doctor --workspace demo-business --coding-user memory-coder
./memory-kit plan --workspace demo-business --coding-user memory-coder
./memory-kit install --workspace demo-business --coding-user memory-coderdoctor and plan are read-only. They check the account boundary, source/tool permissions, runtime versions, root and required listener availability. Resolve failures before installation. They never silently install dependencies or change an existing service.
These commands check different parts of setup; a successful earlier check does not certify the later steps:
| Command and account | Prerequisites checked | Still needs verification |
|---|---|---|
doctor / plan, operator |
Separate standard coding account, protected source and executables, Node 24, PostgreSQL 18 initdb/pg_ctl/psql/postgres, installation root, service port and unlocked default Keychain |
Coding project compatibility, SSH transfer, per-item Keychain access, backup/restore tools and live client behavior |
connect-check, coding account over the intended transfer route |
Explicit expected UID, standard-account membership, project and managed configuration, no binding lock, and unlocked default Keychain | Credential creation/storage, service availability and actual client loading |
connect / reconnect, coding account |
Project binding and secure-store write for the supplied token; reconnect keeps the existing workspace and URL | Server acceptance of the token and current server authorization |
run, coding account |
Intact binding, supported instruction context and readable agent credential | MCP availability, token validity, native client trust/configuration and successful recall/submission |
status, operator |
LaunchAgent loaded state, PostgreSQL running/stopped/unknown state and start-at-login configuration | MCP health, credentials and business operations |
The operator's Keychain must already be unlocked and readable/writable for doctor, plan and install. Preflight does not unlock it or prove an individual credential item's access permissions. PostgreSQL backup and restore tools are listed above for the remaining recovery acceptance; doctor does not validate pg_dump or pg_restore.
install creates a dedicated PostgreSQL cluster, role credentials, the business workspace and a verified private software copy. Credentials are generated in memory and stored in Keychain; no manual SQL or password argument is required. Installation does not start the background service.
The default root is ~/Library/Application Support/Agent Memory Kit/demo-business. It holds private configuration, database storage and versioned software outside the checkout. Do not share it with the coding account or publish it. If you choose a different --root, use that same root for all subsequent commands.
macOS limits Unix socket paths. Setup requires the encoded path <root>/s/.s.PGSQL.47893 to use fewer than 104 bytes, leaving room for its terminating NUL. A long account home or workspace can make the default root too long. Choose a shorter owner-private absolute root outside the checkout, then repeat doctor, plan and install with that same --root; do not move an existing database or loosen directory permissions.
An identical reinstall preserves the instance, credentials and database. Changed software refuses in-place replacement until the protected upgrade procedure is supported. Incomplete database initialization needs operator recovery; do not delete database files to bypass an error.
./memory-kit start --root "$HOME/Library/Application Support/Agent Memory Kit/demo-business"
./memory-kit status --root "$HOME/Library/Application Support/Agent Memory Kit/demo-business"The service exposes http://127.0.0.1:47892/mcp. PostgreSQL uses a Unix socket under the owner-private root, with no TCP database listener. Coding sessions receive only restricted memory API access.
start loads the owner LaunchAgent and enables start at login. A loaded job alone is not a healthy-service check. status distinguishes loaded service, running PostgreSQL and start-at-login configuration. stop drains the owned service/database and removes start-at-login configuration. Run start explicitly to enable it again; unrelated services must remain untouched.
After a Python package update, inspect autostart_python_missing in status and run start again if the recorded executable disappeared. If its configured Python selector is also gone and start refuses, run stop, then repeat the same candidate's install command under the available supported Python to refresh that selector. Then run start and confirm with status. An unknown PostgreSQL state is not reported as stopped: restore the required PostgreSQL 18 tools before attempting recovery.
The launch-context approval decision is complete. The current candidate supports
exactly claude-code/2.1.318 and claude-code/2.1.319: 2.1.319 observer candidate.
2.1.318 is retained only for compatibility, not recommended; 2.1.319 includes
security fixes. Changelog review and binary string comparison are triage, not
proof of identical loading behaviour; discovery coverage remains unproven.
Rollout remains NO GO with the native rehearsal gate unchanged. The real
two-UID boundary is untested because the rehearsal is not being run (James,
9 November 2026). See the whole-candidate repairs and open findings. See the
installation plan.
Host bootstrap of the fixed root-owned Context Authority and delegated protected
software is a separate, unrun gate. Existing service installation does not create
that authority. Never expose the private service root or change native settings
to satisfy the observer.
Once that separately authorised host setup exists, use connect-check --review-context --native-argv '<JSON argv>' over the existing operator transport.
It emits canonical current and prospective bytes for operator --root <SERVICE_ROOT> -- context approve on stdin; approval re-hashes the display and
requires typing the full digest at /dev/tty. No --yes exists. The --root
argument retains the existing operator CLI shape; authority discovery always uses
the fixed path, not that argument. operator --root <SERVICE_ROOT> -- context revoke requires no coding-side packet. Registry revocation does not stop a
running client or revoke its server token.
Pass exactly the same --native-argv JSON to preflight, connect and reconnect;
run receives those same native arguments after --. For example, an already
chosen model/permission command can be represented as
["claude","--model","existing-model","--dangerously-skip-permissions"].
The following transfer examples are schematic until that reviewed argv and host
setup are supplied. Unsupported clients/flags/settings refuse; they are never
silently redirected to another model or configuration directory.
The agent receives a read_submit token. It can read published facts and propose pending facts, but cannot publish or obtain database credentials. From the operator terminal, first create a fictional business member:
./memory-kit operator --root "$HOME/Library/Application Support/Agent Memory Kit/demo-business" -- \
members add --workspace demo-business --user coding-agent --role memberToken output must go directly through a secure input pipe to memory-kit connect in the coding account's own credential context. Never display it, put it in an argument, copy it to chat or save it in an ordinary file. connect reads stdin and stores the token in that account's Keychain; TTY input is refused. At launch, run passes the plaintext token in the native child's AGENT_MEMORY_TOKEN environment variable; its subprocesses inherit it. Tool output can persist it in plaintext transcripts. The token must remain limited to read plus pending-submit for its single workspace; this wrapper cannot prevent the child from disclosing it.
The intended cross-account route is a pre-authorized SSH pipe into the coding account's active, unlocked session. Follow the preflight, transfer and re-key commands. This requires separately prepared SSH access and verified host identity. Setup does not enable Remote Login, bypass authentication or assume sudo unlocks another user's Keychain. The transfer has not passed two-account acceptance. Do not improvise a file or terminal-display workaround. The final novice walkthrough remains a release gate until that exact route is tested.
The destination command has this shape; its stdin must be the secure token stream:
./memory-kit connect --project "$HOME/Projects/business" \
--workspace demo-business --url http://127.0.0.1:47892/mcp \
--instance demo-business-agent \
--native-argv '["claude"]'Run it from the coding account through the delegated protected software copy. It writes .agent-memory.json, managed project instructions and MCP entries in .mcp.json and .codex/config.toml. Configuration contains a Keychain reference and ${AGENT_MEMORY_TOKEN}, never the token. Unrelated configuration is preserved; ambiguous or conflicting bindings are refused.
Use the same coding-client configuration context for the SSH preflight, token transfer, reconnection and later run. In particular, existing CODEX_HOME and CLAUDE_CONFIG_DIR selectors must name the same absolute directories in each context. An interactive launcher may set values that an SSH command does not inherit. Keep those existing settings intact and verify the intended context before minting a token; do not select an empty configuration directory just to bypass a refusal.
The supported local subset can review existing global instruction and personal
logthis sources. It inventories complete supported trees and absences. Imports,
plugins, hooks/frontmatter, active managed preferences/policy, linked worktrees,
unknown settings and unknown context selectors refuse conservatively. Auto-memory
must already be disabled in observed settings. Keep incompatible sources intact;
do not select an empty native home or edit settings to manufacture approval.
Business instructions reuse the existing log this intent. They do not install a competing capture command or a second memory store. Native model and authentication settings remain in place; this is not certification of every existing instruction setup.
Follow the existing tools' search and verification guidance for paraphrased questions and missing knowledge.
From the coding account, launch the installed client through the profile:
./memory-kit run --profile "$HOME/Projects/business/.agent-memory.json" -- claudeThe wrapper validates the binding, retrieves only the agent token and passes it to that child. It does not change model routing. Restart the client after configuration changes. Only the reviewed exact Claude release/subset can launch; other clients and wrappers remain unsupported. The chosen path still needs native acceptance.
Any inventoried local edit requires a fresh operator review; managed-region consistency alone is insufficient. Changes to managed memory configuration fail closed. Rotation requires revoking the old server token and explicitly reconnecting to a fresh Keychain instance. Do not overwrite secure-store entries or change workspace implicitly.
Use the installed-client acceptance charter.
-
Ask a business question in the coding session and verify the empty state.
-
Explicitly select business capture through the existing
log thisentry point to propose a sourced fictional fact. It must report “submitted for review” and remain absent from reads. -
From the operator terminal, list proposals and review the returned proposal ID:
./memory-kit operator --root "$HOME/Library/Application Support/Agent Memory Kit/demo-business" -- \ proposals list --workspace demo-business ./memory-kit operator --root "$HOME/Library/Application Support/Agent Memory Kit/demo-business" -- \ proposals review --workspace demo-business --submission PROPOSAL_ID
-
Read the complete content and confirm the exact displayed hash in the protected terminal. Restart the service and recall the fact from a fresh coding session with its source.
-
Withdraw the knowledge, revoke the token and test outage separately. The client must respect each change and never fall back to personal memory.
The protected operator also has knowledge list, knowledge revoke, tokens list, tokens revoke and membership commands. These remain human administration actions. The coding account must not obtain that authority.
When you have finished testing, or before maintenance, stop the service:
./memory-kit stop --root "$HOME/Library/Application Support/Agent Memory Kit/demo-business"./memory-kit uninstall --root "$HOME/Library/Application Support/Agent Memory Kit/demo-business"Uninstall stops owned services and removes verified software while preserving database, configuration and Keychain entries. Removing a checkout does not delete knowledge. The identical candidate can reinstall its software without reinitializing the database.
Protected backup, current-authority recovery and upgrades remain incomplete release gates. The bundled restore utility produces a quarantined offline archive. That is not an activated recovered service: an old snapshot may predate withdrawals, revoked access and review decisions. Do not activate it as a shortcut.
The installation plan defines scope and installer acceptance defines the required output. Local runtime and native component checks do not replace a clean two-account client installation. The previous file-vault path remains the public default until all required checks and independent review pass.