FAQ

Short answers to questions that come up before reading the docs in depth. For symptom-driven help see Troubleshooting.

---

Do I need an LLM API key?

No, for ingest, vector search, CRUD, and the Web Console. Yes, for consolidation, conflict resolution, importance scoring, and tag extraction. The on/off gate is llm_model — until it is set, POST /api/v1/admin/consolidate runs but processes zero memories. Turn it on with hebb config set llm_model openai/gpt-4o-mini; a hosted provider also needs hebb config set llm_api_key sk-... (a local/proxy model does not).

Which LLMs are supported?

Anything LiteLLM supports — OpenAI, Anthropic, Google, Azure, Bedrock, Qwen, GLM, Kimi, Ollama, vLLM, and many others. Switch with hebb config set llm_model <provider>/<model> and, for non-OpenAI endpoints, llm_base_url. See Multi-model.

Where is my data stored?

In a SQLite file plus a JSON knowledge graph in your workspace directory. Run hebb config get workspace to see the resolved path. Default order: $HEBB_HOME → nearest hebb.json → ~/.hebb/.

Can I use this in production?

v0.1.x is labeled experimental. CRUD and search are stable; consolidation and forgetting policies may change in minor releases. The HTTP server has no built-in auth and CORS is wide open by default — put it behind a reverse proxy with auth before exposing it. PostgreSQL is supported for higher concurrency.

How does it compare to mem0?

Both store agent memories with embeddings. Hebb Mind differs by (a) running as a single local binary with a built-in Web Console and MCP server, (b) implementing a neuroscience-inspired consolidation cycle (encoding → consolidation → forgetting), and (c) defaulting to fully local inference. mem0 has more managed-cloud polish; we have more local control. See Migration from mem0.

Is the Web Console multi-tenant?

No. The Console (and the underlying API) currently operates on a single workspace at a time, with no auth and no per-user isolation. Use partitions (partition_id) to separate logical buckets of memory inside one instance, but treat them as namespaces, not security boundaries.

How do I export my memories?

The HTTP API can dump them all:

curl "http://localhost:8321/api/v1/memories?limit=200&offset=0" > memories.json

For a full backup, copy the workspace files (see "How do I back up the database?" below). A first-class hebb export command is on the roadmap.

How do I back up the database?

Stop the service first to avoid copying a half-written WAL, then copy the workspace:

hebb service stop
cp -a "$(hebb config get workspace)" ~/backup/hebb-$(date +%Y%m%d)
hebb service start

For PostgreSQL, use pg_dump against the URL set in pg_url.

Can I run it on PostgreSQL?

Yes:

hebb config set storage_type postgresql
hebb config set pg_url "postgresql://user:pass@host:5432/hebb"

The schema is created automatically on first start. See Storage Backends for pgvector requirements.

Is there a hosted version?

Not currently. Hebb Mind is designed to run locally or on your own infrastructure. If a hosted option appears in the future it will be announced on the GitHub repo.

How do I deploy with Docker?

A reference image and docker-compose.yml (with PostgreSQL + pgvector) live in Storage Backends → Docker. The short version:

docker run -p 8321:8321 -v ~/.hebb:/data ghcr.io/afx-team/hebb-mind:latest

What's the license?

MIT. See LICENSE in the repo. You can use it commercially, modify it, redistribute it — just keep the copyright notice.

How do I contribute?

Issues, PRs, and discussion are welcome at https://github.com/afx-team/hebb-mind. The most useful contributions right now are: real-world bug reports with hebb doctor output, integration recipes (Cursor / Continue / Cline), and benchmark reproductions on hardware we don't own. See CONTRIBUTING.md in the repo root.

Why "Hebb Mind"?

Hebb Mind is named after Donald O. Hebb (1904–1985), the Canadian neuropsychologist whose 1949 book The Organization of Behavior gave us Hebbian learning — "neurons that fire together, wire together." That rule is the heart of the framework: the tag knowledge graph wires co-occurring concepts together and strengthens the link each time they recur, while consolidation keeps what is reinforced and forgetting prunes the rest.

The hippocampus — the project's original name — lives on inside Hebb Mind as the name of the working-memory partition (mem_hippocampus): the inbox where new memories land before consolidation, mirroring the brain region that gates new experience into long-term memory. See Memory Lifecycle.

Is the embedding model multilingual?

It depends on what hebb setup selected. By default (--profile default), --language en selects a small English-only model (all-MiniLM-L6-v2, dim 384, ~90MB) and --language zh / --language multi selects intfloat/multilingual-e5-small (multilingual, dim 384, ~470MB). The high-quality bge tier — BAAI/bge-large-en-v1.5 (English) / BAAI/bge-m3 (multilingual), both dim 1024, 1–2GB — is opt-in via hebb setup --profile best. To swap models after install — including the dimension-change → re-embed flow — see Switch the Embedding Model.