Architecture

Memory Layers

GCTRL organises memory into four parallel tiers. Each tier is a different representation of the same knowledge, optimised for a different kind of read. Agents pull from the tier that fits the question - and write conclusions back - so memory compounds instead of going stale.

Why four tiers

A single store cannot serve every need. An authoritative one-line answer, the raw evidence behind it, the structural provenance, and a human-readable explanation are four different things. GCTRL keeps all four in sync and exposes each one directly.

TierNameBacking storeBest read for
HOTEntity DossiersPostgres / graphAuthoritative summary - answer with confidence
WARMText ChunksQdrant (+ source)Evidence passages for grounded answers
COLDKnowledge GraphNeo4jStructure, relationships, provenance
WIKICurated ProseWiki viewHuman-readable explanation

HOT - Entity Dossiers

Compiled, authoritative summaries for important entities, distilled by FUSE.

  • Pinnable - operators can pin a dossier so it is never decayed or evicted.
  • Trust-scored - each dossier carries a trust signal so consumers know how reliable it is.
  • Direct answer. When an agent reads a dossier, it has the canonical view - it should state the fact, not hedge.

WARM - Text Chunks

The source passages plus their embeddings, written by KEX into the vector store.

  • Evidence layer - these are the actual snippets that back an answer and supply citations.
  • Heat-tracked - usage is measured; frequently retrieved chunks run "hotter."
  • Used in answers - Talk-to-Graph retrieves chunks here for grounded, cited responses.

COLD - Knowledge Graph

The structured entities and relationships in Neo4j.

  • Permanent - the durable backbone of all knowledge.
  • Clearance-gated - traversal and reads respect access control.
  • Provenance - the graph is where you trace how facts connect and where they came from.

WIKI - Curated Prose

LLM-distilled, cross-linked human-readable pages built by FUSE over the graph.

  • Readable - narrative prose rather than triples or vectors.
  • Cross-linked - pages reference related entities for navigation.
  • For explanation - the right layer when a human (or agent) needs context, not just a fact.

Memory governance

Tiers stay healthy through an automatic governance cycle that runs continuously and can also be triggered on demand:

  • Decay - stale, unused items lose heat and are aged out.
  • Dedup - near-duplicate chunks are collapsed to keep the warm tier lean.
  • Promote - frequently-used warm → hot: chunks that keep proving useful get compiled into dossiers.
  • Evict - distrusted or cold items are removed so noise does not accumulate.

Health snapshot

A /memory/health snapshot reports the state of the tiers:

MetricMeaning
CoverageHow much of the corpus is represented across tiers
Store sizesItem counts per tier (hot / warm / cold / wiki)
Heat distributionHow retrieval is spread across chunks
Trust distributionSpread of trust scores across dossiers

On-demand maintenance

Operators (and agents with the right scope) can trigger a maintenance run to execute the decay/dedup/promote/evict cycle immediately rather than waiting for the scheduled pass.

The payoff: compounding memory

Agents read the right layer for the job:

  • Authoritative dossier (HOT) → state the answer.
  • Evidence chunks (WARM) → ground and cite.
  • Graph (COLD) → trace provenance and structure.
  • Curated prose (WIKI) → explain in context.

And they write conclusions back after each task. Because governance keeps the tiers clean and promotes what proves useful, every session leaves the memory richer than it found it.

See also

Architecture · Modules · Agents & MCP · GCTRL Memory Skill