Lewati ke konten utama

FAQ

Where is HMG's data stored?​

In the store directory on your machine, by default <user data dir>/hmg/stores/default — on macOS / Linux usually ~/.local/share/hmg/stores/default. Use --store <path> to override, or hmg doctor --verbose to see the actual path. Data stays local by default.

The store contains:

  • long-term memories
  • indexes
  • observations
  • local runtime state

Default location and layout: see Configuration.

How should I think about local data and privacy?​

HMG is local-first by default:

  • memories, indexes, and observations all live in the local store
  • you decide where the store directory goes
  • you can back up, migrate, isolate, or delete the store directly
  • when several agents share one store, data boundaries come from scope, governance state, and local directories together

Two rules:

  1. Don't put secrets, tokens, passwords, or sensitive customer data into regular memories.
  2. For credentials that must persist, use hmg secret store — never plaintext in regular memories.

Do I need to log in? How do plans work?​

No forced login — basic memory features work without it. To unlock developer-edition features and higher memory capacity:

  • Upgrade your plan in the user center on the website, then run hmg login locally
  • Login requires network: HMG verifies the account with the remote service, identifies the plan, receives a signature, and unlocks after local verification
  • Login performs no migration and modifies no local memory data: local reads/writes always use the OS username as tenant; the account is only for identity mapping in cloud sync
  • Logging out or switching accounts only clears the linked-account field in the store config; local memories are untouched

How is HMG different from chat history?​

Chat history is a complete running log of conversations, piled up chronologically — hard to search. HMG memory is information you'll use in the future: decisions, root causes, constraints, preferences, verification results — structured, searchable, correctable, scoped. HMG is not "remembering what you said" but "remembering what stays useful".

How is HMG different from a vector database?​

A vector database only does semantic similarity search. On top of that HMG provides:

  • Scope (tenant/workspace/repository/branch) — context isolation
  • Multi-route retrieval (semantics + keywords + graph relations, etc.)
  • Correction and governance (information goes stale: correct, demote, quarantine, tombstone — with audit kept)
  • Structured queries (decision tracing, risk lists, impact analysis, knowledge graph exploration)

In short: a vector DB is "store + search"; HMG is "memory that expires, gets corrected, and carries context".

Do I need to write memories manually?​

Mostly no. With an agent integrated, memory writing is fully autonomous: the agent calls memorize incrementally as decisions and exchanges happen, writes a handoff at task end, and corrects stale information with correct. Manual writing fits what you want to deliberately deposit: project conventions, stable preferences, long-term constraints.

Does the agent need to pass scope when calling HMG?​

No — and it shouldn't. In agent integration, scope is mechanically inferred from the session directory and injected by hooks (all four levels: tenant / workspace / repository / branch); wrong values passed by the agent are silently corrected. Only manual CLI use may involve --scope. See Scope.

Why can't project sessions see memories from ephemeral sessions?​

That's scope isolation by design. Ephemeral sessions not attached to a project (e.g. Codex desktop Chats) all share the personal/chats/main scope, isolated from project scopes. If a user preference stored in an ephemeral session is also needed in a project, memorize it again inside the project session.

Won't memories grow endless and messy?​

HMG has multiple layers of denoising:

  • Write-source control: in agent integration, durable memories only enter through the two active channels memorize / handoff — no mechanical capture of raw conversations or command output
  • Correction / demotion / governance: stale info can be corrected or demoted; sensitive info governed
  • Exact dedup: identical content is never stored twice
  • store hygiene: hmg store hygiene cleans orphan edges/indexes
  • Scope isolation: projects/branches never pollute each other

Combined with the writing guidance in Best Practices, memories stay concise.

What workflows is HMG suited for?​

Any workflow where the agent should keep remembering context. Typical:

  • long-term collaboration on solo projects
  • multi-session, multi-day tasks with continuous handoff
  • experience accumulation across repos and branches
  • several agents on one machine sharing long-term memory

Can I migrate to another computer?​

Yes. Two ways:

bash
# 1. Copy the store directory to the new machine
# 2. Official migration command
hmg store migrate --from /old/store --to /new/store --backup --apply

The simplest backup is copying the whole store directory.

Or export first with hmg export --format json and import on the new machine. See Backing up or migrating local data.

How should sensitive information be handled?​

  • Don't write API keys, tokens, or passwords into regular memories.
  • Credentials that must persist: use the secret vault — hmg secret store <name> <value>.
  • Reveal on demand: hmg secret reveal <name>.
  • At write time HMG auto-redacts structured sensitive info (connection strings, password=xxx, Bearer tokens, private key blocks); natural-language sensitive content is not auto-detected.
  • Sensitive info written by mistake? Govern immediately: hmg govern <atom-id> --action tombstone --destroy-payload --reason "sensitive info written by mistake".

Governance options: see Removing or isolating sensitive information.

Can I delete memories?​

Yes, but HMG recommends governance over physical deletion to keep the audit trail:

bash
# Tombstone (logical deletion, payload optionally destroyed)
hmg govern <atom-id> --action tombstone --destroy-payload --reason "no longer needed"

# Quarantine (kept, hidden from recall)
hmg govern <atom-id> --action quarantine --reason "not needed for now"

# Demote (may no longer be needed)
hmg correct <atom-id> --action demote-possible --reason "no longer applicable"

Observations can be removed with hmg obs forget.

Can I import existing conversation history and memory files?​

Yes. Say "import memories into HMG" to your agent (the hmg-memory-import skill): it scans conversation history and host-native memory files, extracts items with long-term value, classifies them by project / personal, and writes them after your confirmation. See Importing existing memories.