hypabase-memory（低碱基记忆）

Hypabase Memory

Persistent memory for agents. Stores preferences, decisions, facts, and events as a connected knowledge graph. Recalled by who, what, when, or why.

Setup

Add the MCP server to your OpenClaw config (~/.openclaw/openclaw.json):

{
  "mcpServers": {
    "hypabase-memory": {
      "command": "uvx",
      "args": ["--from", "hypabase", "hypabase-memory"],
      "env": { "HYPABASE_DB_PATH": "hypabase.db" }
    }
  }
}

uvx handles all Python dependencies automatically. Requires uv.

Quick start

Store a memory — one verb, roles for the participants:

remember(penman='(prefers :subject Alice :object Python :memory_type semantic)')

Recall it:

recall(entity="Alice")                              # everything about Alice
recall(entity="Alice", action="assign", role="subject")  # what Alice assigned
recall(entity=["Alice", "Bob"])                     # memories involving both
recall(mood="planned")                              # all plans

When to Remember

Store a memory when the user:

• Makes a decision or states a preference
• Shares a fact about themselves, their team, or a project
• Assigns a task or delegates work
• Describes an event, meeting, or outcome
• Explains a procedure or workflow

PENMAN Notation

Every memory is a verb with participants in role slots:

(verb :role "entity" :role "entity" ...)

Examples:

(prefers :subject Alice :object Python :memory_type semantic)

(assigned :subject Alice :object "billing task" :recipient Bob
 :instrument Jira :locus Monday :tense past :memory_type episodic)

(has :subject "quick sort" :attribute "time complexity"
 :value "O(n log n)" :memory_type semantic)

Multiple atoms in a single call:

(deployed :subject Alice :object API :locus Monday :tense past)
(reviewed :subject Bob :object API :locus Tuesday :tense past)

One action per memory. When a sentence contains multiple actions, decompose into separate atoms. Shared entities link them in the graph.

"Alice told Bob to migrate the database":

(told :subject Alice :recipient Bob :object "database migration" :tense past)
(migrate :subject Bob :object database :mood planned)

Any role slot can hold a nested atom:

(believes :subject Alice :object (is :subject deadline :value Friday))

Entity Naming

Same string after lowercasing = same entity. Different strings = different entities until consolidate() merges them.

• Pick one canonical name per entity and reuse it.
• Use full descriptive names: "machine learning" not "ML", "JavaScript" not "JS".
• Call consolidate() periodically to merge similar names via semantic similarity.

Tools

remember(penman, source?, confidence?)

Store memories as PENMAN atoms.

• penman (required): One or more PENMAN atoms.
• source (optional): Provenance source. Default: "memory".
• confidence (optional): 0.0–1.0. Default: 1.0.

recall(entity?, action?, role?, memory_type?, mood?, negated?, since?, before?, limit?, min_strength?)

Query memories. At least one parameter required.

• entity: Name or list of names.
• action: Filter by verb.
• role: Filter by role (subject/object/instrument/recipient/source/locus/attribute/value).
• memory_type: episodic / semantic / procedural.
• mood: actual / planned / uncertain / normative / conditional.
• negated: true = only negated, false = only positive.
• since / before: ISO date strings.
• limit: Max results (default: 10).
• min_strength: Minimum memory strength threshold.

consolidate(entity?)

Merge similar entities and compress repeated memories.

forget(older_than_days?, min_strength?, entity?)

Expire old or low-strength memories (soft delete).

Reference

Roles

Eight roles. Fill in what applies, skip what doesn't.

Memory types

Moods

Modifiers

Environment variables

• HYPABASE_DB_PATH -- SQLite database path (default: hypabase.db)
• HYPABASE_EMBEDDER -- Embedder for semantic search:

- fastembed (default) -- BAAI/bge-small-en-v1.5 via ONNX - openai -- text-embedding-3-small (requires OPENAI_API_KEY) - sentence-transformers -- all-MiniLM-L6-v2 - none -- disable embeddings