How Lambo Memory Works

Mooshik links Lambo directly in-process via the lambo Rust crate.

Instead of treating memory as a flat text log, Lambo maintains an in-memory directed property graph backed by asynchronous write-behind storage.

Concept Types

Lambo categorizes extracted knowledge into five distinct concept types:

1. entity: Named components, services, tools, databases, and workspace systems.
2. logic: Invariants, algorithms, transformation steps, and procedural rules.
3. constraint: Technical boundaries, port allocations, architectural limits, and version caps.
4. resource: File paths, documents, git commit references, URLs, and screenshots.
5. observation: Empirical outcomes, test results, benchmarks, and performance measurements.

Different concept types decay at different rates. Constraints resist eviction longer than passing observations.

Graph Relations

Edges in the graph express structural relationships between concepts:

• parent_of: Represents hierarchical ownership and component containment.
• derives: Connects source documents or observations to derived concepts.
• depends_on: Tracks operational dependencies between components.
• constrains: Attaches constraints and boundaries to target entities.

Companion Memory Tools

Mooshik exposes three in-process memory tools to the companion model:

Tool	Purpose	Output
lambo_recall	Queries the graph using semantic similarity and graph traversal.	Returns ranked concepts with relevance scores and blast radius.
lambo_derive	Writes new concepts and edges into the in-memory graph.	Returns confirmation of created or updated concept IDs.
lambo_stats	Inspects graph size, node counts, and durability metrics.	Returns node counts, edge counts, and flush queue depth.

The Memory Lifecycle

flowchart LR
    Ingest["1. Workspace Event<br/>(Watcher / Ingester / Turn)"] --> Derive["2. lambo_derive<br/>(WriteLane Serialized)"]
    Derive --> Graph["3. In-Memory Graph<br/>(Immediate Read Consistency)"]
    Graph --> Flush["4. Write-Behind Flush<br/>(Postgres / SQLite)"]
    Graph --> Daemon["5. Daemon Rescoring<br/>(Earned Canonization)"]
    Graph --> Reflect["6. mooshik reflect<br/>(Prose Synthesis & Merging)"]

1. Extraction: The watcher, conversational turn, or MCP server extracts concepts from activity.
2. Derivation: Concepts are written to memory through lambo_derive, serialized by WriteLane.
3. Storage: The in-memory graph updates immediately, followed by asynchronous write-behind flushes.
4. Rescoring: The background daemon computes structural scores across time.
5. Consolidation: mooshik reflect merges paraphrase twins and synthesizes timeline prose.