Agent memory & knowledge graphs (MCP)

LessDB ships an embedded, durable memory tier for agents: linked context notes, RAM tables with point lookups, and an openCypher subset — all exposed through the same MCP server as the SQL database. One process replaces the "context vault + graph DB + vector store" stack.

Start the MCP server

// claude_desktop_config.json / codex config — point at the less binary
{ "mcpServers": { "lessdb": { "command": "less", "args": ["mcp", "--dir", "~/.less"] } } }

Tools your agent gets

group	tools	what they do
database	`less_query`, `less_explain`, `less_schema`, `less_stats`, `less_tables`, `less_optimize`	full SQL over MergeTree/SharedMergeTree tables
context	`context_put`, `context_find`, `context_link`, `context_neighbors`, `context_path`	titled/tagged notes, typed links, BFS neighborhoods, shortest paths
memory	`memory_create`, `memory_insert`, `memory_sql`, `memory_get`, `memory_compact`	Arrow-backed RAM tables with a hash PK index — O(1) latest-row lookups
graph	`less_cypher`	`MATCH` with variable-length patterns, `WHERE`, aggregation, `CREATE`/`DELETE`/`SET`

Example session

agent: less_query  →  SELECT count(*) FROM incidents WHERE status='open'
agent: context_put  →  key "incident/42", title "DB failover", tags ["incident","postmortem"]
agent: context_link →  "incident/42" -[follows_up]-> "incident/39"
agent: context_path →  "incident/39" → "incident/42"          (shortest path)
agent: memory_create →  sessions(id Int64, user Utf8, ts Timestamp)
agent: memory_get   →  key: sessions/id = 7                    (O(1), latest row wins)
agent: less_cypher  →  MATCH (a:incident)-[:depends_on*1..3]->(b) RETURN a.id, b.id

Properties that make it agent-safe

Durable + crash-safe: every mutation snapshots atomically

(tmp + rename); agents can be killed and resume.

Tenant-scoped: less mcp --tenant <name> isolates each agent's

contexts/memory/vectors under <data_dir>/tenants/<name>/ — SQL tools stay on the shared engine.

Same data, two front doors: humans get SQL/TUI; agents get tools —

over the same engine, so agent findings are immediately queryable by people and vice versa.

Full API in Getting started and the architecture doc's in-memory tier section.