← All work

Planets

Knowledge infrastructure — structured claims that survive the agent

Last devlog: June 25, 2026

EPHEMERAL — AGENTS PERSISTENT — THE PLANET INTEGRATION LAYER — THE MEMBRANE CrucibleObserve · learn · decideephemeral · deterministicCrucibleanother agentephemeralIntegrationLayerQuality gatesDedup · state derivationdeterministicClaim StoreClaims · provenance · edges+ pgvector embeddingsPostgres · persistentQuery ServerSQL · vector · graphthree modalities, one storeEmbeddingServiceprobabilistic · shared infraAgentquerydeltasdeltasclaims + edgesqueryembed (write)embed (query)

Key boundary: The line between ephemeral agents and persistent knowledge. Agents come and go; claims persist. The Integration Layer is the membrane — it applies quality gates and provenance requirements before anything enters the store. Without this boundary, knowledge dies with the agent that produced it.

Notable decision: Three retrieval modalities (deterministic SQL, semantic vector search, graph traversal) in one Postgres database — no external vector store, no external graph engine. The routing happens inside the query server against local SQL, not between external services.

Planets inverts the standard multi-agent framing. In most systems, agents are primary — they have capabilities, goals, and state, and the data they produce is a byproduct. In Planets, the data model is primary. The accumulated state — the Planet — is the thing that exists and persists. Agents are subservient to it; they exist to enrich it, maintain it, and serve its coherence.

The architecture has three sides. Ephemeral agents (Crucibles) observe a system, learn a fact, or make a decision, and emit structured deltas with provenance. An integration layer receives those deltas, runs quality gates, resolves duplicates, derives state claims, and persists to a claim store — a single Postgres database with pgvector for semantic search and typed edges for graph traversal. A query server gives agents and the operator one interface with three retrieval modalities: deterministic SQL, semantic vector search, and graph traversal, all against the same store.

The load-bearing boundary is between the ephemeral agent and the persistent knowledge. Agents come and go; claims persist. The integration layer is the membrane — it applies quality gates and provenance requirements before anything enters the store. Without this boundary, knowledge dies with the agent that produced it. The planet is empty every time an agent terminates. The boundary is what makes knowledge accumulate.

The first Planet is Silent Forest — structured knowledge about infrastructure operations, populated passively from live monitoring telemetry. The claim model is general: it holds assertions with provenance, which means it holds any knowledge an agent can justify. The architecture is not limited to operational telemetry; it's limited to what agents can observe and defend.

Devlog posts about Planets

Two red teams and a spike: designing agent memory that survives the agent

An agent memory architecture proposal that survived two adversarial red team passes and a humanization spike before a single line of implementation. The methodology caught four critical design flaws — including one that would have made semantic search silently useless.