Lambo

Lambo is agentic graph memory for multi-agent AI operations. It stores what your agents learn, recalls relevant memory on request, and coordinates edits across many small agents that never talk to each other directly.

A single lambo serve process owns each session. Agents connect to that process and share its in-memory graph. A background daemon consolidates what they write.

flowchart LR
  subgraph Surfaces["Agent surfaces"]
      MCP["MCP clients<br/>(agents)"]
      CLI["Command line verbs"]
      WEB["lambo serve-web<br/>(browser, read-only)"]
  end
  subgraph Writer["lambo serve (single writer)"]
      GRAPH["In-memory graph<br/>concepts and edges"]
      DAEMON["Background daemon"]
      CANON["Canonization"]
      EMB["Embedder"]
  end
  subgraph Durable["Durable store"]
      DB[("memory / sqlite /<br/>cockroach")]
  end
  MCP --> GRAPH
  CLI --> GRAPH
  WEB -. "read-only" .-> DB
  GRAPH --> EMB
  GRAPH --> DAEMON
  DAEMON --> CANON
  GRAPH -. "write-behind flush" .-> DB

The project narrative

Multi-agent coding produces a flood of observations. Each small agent sees one slice of the work, writes it down in its own words, and moves on. That memory lives in tool call history and dies when the context window closes.

That is the problem Lambo was built for, and Origin is the author’s account of how the idea arrived: what working with coding agents made obvious, and where a promotion model borrowed from veneration and canonization came from.

Lambo gives those agents a shared memory that outlives the context. It treats each session as a graph of concepts. A concept is something your agents learn, such as user schema, auth middleware, or retry budget is 3. Typed edges connect concepts that depend on, produce, or relate to each other.

Three properties make this memory useful at scale.

• It consolidates. Many agents write the same idea in different words. Lambo matches the duplicates, deduplicates them, and lets recall return one canonical fact instead of fifty near-identical notes.
• It promotes by evidence. Importance is earned, not declared. The background daemon scores concepts from structural evidence such as peer count, blast radius, and how long edges have lived. A concept that accumulates evidence earns canonical status. Canonization has the arithmetic, and Evidence and evaluation has the transcripts that verify it from a plain SQL client.
• It retracts safely. Each concept carries a blast radius, so an agent knows how many things depend on what it is about to change. Coordination tools such as lambo_reserve and lambo_inspect turn that knowledge into action.

Durability is write-behind and eventual. Mutations batch into a few round trips to a durable store, so a burst of writes drains quickly and the session survives restarts.

Explore

Four ways in. The sidebar has everything else.

QuickstartInstall the binary and run your first session in about 30 seconds.

DemoWatch two agents build one REST API and share load-bearing memory.

CanonizationHow a concept earns canonical status: the score, the three gates, the constants.

Evidence & evaluationLive-cluster measurements, determinism proofs, and the raw transcripts behind every claim.