walrus-memory

Graph-based memory extension for OpenWalrus agents. Runs as an extension service managed by the daemon.

Install

walrus hub install openwalrus/memory

Or build from source:

cargo install walrus-memory

Configuration

Installed automatically by walrus hub install. To customize, edit walrus.toml:

[services.memory]
kind = "extension"
crate = "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)

Architecture

Storage: LanceDB with three tables — entities, relations, journals. Embeddings via candle (all-MiniLM-L6-v2, 384-dim). Shared global graph across all agents.

Agent-facing tool

recall — batch semantic search + 1-hop graph traversal. Agent sees this as a regular tool. Automatically called before each agent run (auto-recall).

Background extraction

After each agent run, the daemon triggers async extraction via the extension's Infer capability. An extraction LLM reviews the conversation and calls two internal tools:

recall — check existing memories for dedup
extract — batch upsert entities and relations

The agent never calls extract directly — it's only used by the extraction LLM.

Lifecycle hooks

Hook	Behavior
BuildAgent	Injects identity/profile entities into system prompt
BeforeRun	Auto-recall: searches memory, injects `<recall>` block
AfterRun	Stores conversation journal, triggers extraction via Infer
Compact	Enriches compaction prompt with recent journals

Development

Normally launched by the daemon. For standalone development:

walrus-memory serve --socket /tmp/memory.sock

License

MIT OR Apache-2.0

walrus-memory 0.0.7