Scoped memory for AI agents on Nostr.

Nomen gives AI agents persistent, searchable, shareable memory using Nostr as the sync layer. Memory is scoped by visibility — from public knowledge to agent-private reasoning — and signed with the agent's keypair. No vendor lock-in, no central database. Your memory travels with your keys.

• Semantic + full-text — HNSW vector similarity + BM25, weighted and composable
• Graph-aware retrieval — Traverses entity connections, contradictions, and provenance chains from search hits (--graph, configurable --hops)
• Entity extraction — Heuristic + LLM-powered via EntityExtractor trait with typed relationships (works_on, collaborates_with, contradicts)
• Cluster fusion — Groups related memories by namespace and synthesizes coherent summaries via LLM
• Confidence decay — Unaccessed memories lose ranking weight over time
• Importance scoring — LLM assigns 1-10 importance at creation

Raw messages flow in from conversations. Consolidation extracts signal from noise:

Raw Messages → grouping → LLM extraction → merge/dedup → Named Memories

• Tier derived from source (DM → personal, group chat → group, self-reflection → internal)
• Checks existing memories by topic + embedding similarity (>0.92), merges instead of duplicating
• LLM flags contradictions, creates contradicts graph edges
• Automatic entity extraction and linking during consolidation

Memories are kind 31234 addressable replaceable events. D-tag encodes visibility, scope, and topic:

{visibility}:{scope}:{topic}

{
  "kind": 31234,
  "content": "{\"summary\":\"Use anyhow for app errors\",\"detail\":\"Prefer anyhow::Result…\"}",
  "tags": [
    ["d", "public::rust-error-handling"],
    ["visibility", "public"],
    ["scope", ""],
    ["model", "anthropic/claude-opus-4-6"],
    ["confidence", "0.92"],
    ["version", "1"],
    ["t", "rust"],
    ["t", "error-handling"]
  ]
}

Indexed visibility and scope tags enable relay-side filtering without prefix matching:

{"kinds": [31234], "#visibility": ["group"], "#scope": ["techteam"]}

D-tag examples:

• public::rust-error-handling
• group:techteam:deployment-process
• circle:a3f8b2c1:shared-notes
• personal:d29fe7c1…:ssh-config
• internal:d29fe7c1…:agent-reasoning

• Memory browser with search
• Consolidation status + "Run Now"
• Pruning controls with dry-run preview
• Entity and graph explorer
• Config viewer + reload

git clone https://github.com/k0sti/nomen.git
cd nomen
cargo build --release

No external database — SurrealDB runs embedded.

# Interactive setup
nomen init

# Store a memory
nomen store "rust/error-handling" \
  --summary "Use anyhow for app errors" \
  --detail "Prefer anyhow::Result for ergonomic error propagation." \
  --tier public --confidence 0.92

# Search (hybrid vector + BM25)
nomen search "error handling"
nomen search "error handling" --graph         # + graph traversal
nomen search "error handling" --graph --hops 2

# Ingest + consolidate
nomen ingest "k0 mentioned switching to Tailscale" --source telegram --sender k0
nomen consolidate

# Entity extraction and cluster fusion
nomen entities --relations
nomen cluster --dry-run

# Start MCP server
nomen serve --stdio
nomen serve --http 127.0.0.1:3000

# Context-VM (Nostr-native)
nomen serve --stdio --context-vm --allowed-npubs <hex-pubkey>

              Nostr Relay (kind 31234)
                      │
                 sync (NIP-42/44)
                      │
                  ┌───┴───┐
                  │ Nomen │
                  └───┬───┘
                      │
                  SurrealDB (embedded)
                ┌─────┼─────┐
                │     │     │
              docs  vectors  graph
              BM25  HNSW    edges
                │     │     │
        ┌───────┴─────┴─────┴───────┐
        │       │       │           │
       CLI   Library  MCP Server  Context-VM

Nostr relay is the source of truth. SurrealDB is a local index. If local state is lost, nomen sync recovers everything.

Every resource-intensive feature can be independently toggled:

relay = "wss://your-relay.example.com"
nsec = "nsec1..."

[embedding]
provider = "openai"                # or "none" to disable
model = "text-embedding-3-small"
api_key_env = "OPENAI_API_KEY"

[memory.consolidation]
enabled = true                     # toggle consolidation
provider = "openrouter"
model = "anthropic/claude-sonnet-4-6"
api_key_env = "OPENROUTER_API_KEY"

[entities]
provider = "openrouter"            # "none" → heuristic fallback
model = "anthropic/claude-sonnet-4-6"
api_key_env = "OPENROUTER_API_KEY"

[memory.cluster]
enabled = true                     # toggle cluster fusion
min_members = 3
namespace_depth = 2

[server]
enabled = true
listen = "127.0.0.1:3000"

[contextvm]
enabled = false                    # toggle Context-VM

src/
├── main.rs         CLI (clap)
├── lib.rs          Public API
├── config.rs       Config loading
├── db.rs           SurrealDB schema, queries, graph edges
├── memory.rs       Memory types, d-tag parsing (v0.1/v0.2)
├── search.rs       Hybrid search + graph expansion + aggregation
├── entities.rs     Entity extraction (heuristic + LLM + composite)
├── cluster.rs      Cluster fusion pipeline
├── consolidate.rs  Consolidation (LLM extraction, merge, dedup)
├── embed.rs        Embedding providers
├── relay.rs        Nostr relay sync (NIP-42, NIP-44)
├── access.rs       Tier enforcement
├── groups.rs       Group hierarchy + scope resolution
├── ingest.rs       Raw message ingestion
├── mcp.rs          MCP server (JSON-RPC stdio/HTTP)
├── cvm.rs          Context-VM (kind 25910, NIP-59)
├── tools.rs        Shared tool dispatch (MCP + CVM)
├── http.rs         HTTP API + dashboard
├── send.rs         Nostr messaging
├── migrate.rs      Import from other formats
└── display.rs      Terminal formatting

web/                Svelte web dashboard
docs/               Specifications

Nomen = Nostr Memory Network.

Extracted from Snowclaw as a standalone memory layer for any Nostr-aware agent.

• Nostr Memory Spec — event format, visibility tiers, relay queries
• Consolidation Spec — pipeline, merge logic, tier derivation
• Architecture — system design, data flow
• Graph & Consolidation Roadmap — feature implementation details

MIT