pipx install + hebb setup + hebb service install. First run downloads a small embedding model (~90MB English / ~470MB multilingual), only if not already present. SQLite + sentence-transformers, zero external services. No API key needed for ingest and hybrid search.
🧠
Conflict-resolving consolidation
An agent merges duplicates and overwrites stale facts — not just append. Bring any LLM via LiteLLM (OpenAI, Claude, Qwen, GLM, Kimi, …).
🔄
Honest forgetting
TTL = base × (1 + log(access)) × importance × exp(-decay × days). Frequently used memories survive; neglected ones decay. Tunable per workspace.
🔍
Three-path hybrid search
Vector + keyword + tag-graph retrieval, scored on recency, importance, and relevance. NetworkX-backed knowledge graph; explore neighbors via the API.
🖥️
Built-in Web Console
Single-page app for memory CRUD, search, partitions, and graph view. Lives at http://localhost:8321/ — no separate deploy.
🔌
REST + MCP + Agent Sync
Claude Code and Codex get automatic recall and turn capture through hooks, explicit memory tools over MCP, and historical session backfill through the Agent Sync hub.
Why "Hebb Mind"?
In 1949, Canadian psychologist Donald O. Hebb (1904–1985) described the rule the brain learns by. Hebb Mind is built on it.
“When an axon of cell A … repeatedly or persistently takes part in firing cell B, … A's efficiency, as one of the cells firing B, is increased.” — D. O. Hebb, The Organization of Behavior (1949) · remembered as “neurons that fire together, wire together.”
Hebb's insight: a memory is not a place — it is a pattern of connection. Concepts that co-occur wire into cell assemblies; a partial cue lights up the rest. Hebb Mind's tag knowledge graph runs exactly that loop.
🔗
Wire
Co-occurring tags gain a graph edge — a cell assembly
→
💪
Strengthen
Each co-activation thickens the edge; consolidation keeps what's reinforced
→
🌐
Complete
Retrieval walks the edges — a cue recalls the whole pattern
→
💨
Prune
What is never co-activated weakens and fades
And the hippocampus — the project's original name? It lives on as the working-memory partition (mem_hippocampus): the inbox every new memory enters before consolidation, just as the brain's hippocampus gates new experience into long-term memory. The brain region became one component; the learning rule became the name.
The Memory Loop
Four stages, in roughly the order the brain runs them — encoding in CA1, replay during slow-wave sleep (Wilson & McNaughton, Science, 1994), pattern-completion in CA3, and the forgetting curve (Ebbinghaus, 1885) doing its quiet work.
pipx install hebb-mind && hebb setup && hebb service install
First run downloads a small embedding model (~90MB English / ~470MB multilingual), only if not already cached — about a minute on the English/fast path, a little longer for multilingual. Want the high-quality bge models (1–2GB)? Use hebb setup --profile best. No pipx yet? One-time setup (brew / apt / dnf / python -m pip). Then open http://localhost:8321/. For LLM consolidation, see the 5-minute path.