Memory

Graph-based persistent memory service — entities, relations, and journals backed by LanceDB.

The memory service (walrus-memory) provides graph-based persistent memory for agents. It runs as a WHS hook service — the daemon routes memory tool calls to it and receives identity data at agent startup.

Configuration

[services.memory]
kind = "hook"
command = "walrus-memory"
enabled = true
config = { entities = ["project"], relations = ["implements"], connections = 30 }

Config field	Type	Description	Default
`entities`	string[]	Additional entity types	`[]`
`relations`	string[]	Additional relation types	`[]`
`connections`	usize	Default limit for graph traversal	`20` (max 100)

Tools

The memory service registers six tools:

Tool	Description
`remember`	Store a typed entity (upserts by composite key)
`recall`	Search entities by full-text query, optionally filtered by type
`relate`	Create a directed edge between two entities
`connections`	Traverse the graph one hop from an entity
`compact`	Trigger context compaction with journal storage
`distill`	Search past journal entries by semantic similarity

See memory concepts for detailed tool schemas and usage.

Capabilities

The memory service declares four WHS capabilities:

Tools — the six memory tools above
BuildAgent — injects identity traits, user profile, and recent journals into the system prompt at agent creation
Compact — enriches the compaction prompt with memory context
EventObserver — observes agent events for knowledge extraction

Entity types

Seven built-in types: fact, preference, person, event, concept, identity, profile. Extend via the entities config field.

Relation types

Seven built-in types: knows, prefers, related_to, caused_by, part_of, depends_on, tagged_with. Extend via the relations config field.

What's next

Memory concepts — storage architecture, tool schemas, identity injection
WHS services — how hook services work
Context compaction — how compact and distill work together