knowledge_graph: no intermediate associative-tag layer — cue→content only, so graph traversal cannot be bounded (MRAgent prerequisite)

[cdeust]

Prerequisite for #361. MRAgent (arXiv 2606.06036) represents memory as a Cue-Tag-Content heterogeneous graph M = (C, G, V) with typed relations R ⊆ C × G × V, where the tag layer is not decoration — it is the mechanism that makes iterative traversal tractable.

Current state (verified 2026-08-06)

grep -niE 'def .*tag|tag_' mcp_server/core/knowledge_graph.py returns nothing. Cortex's graph is cue→content: core/knowledge_graph.py extracts entities and relationships, and core/spreading_activation.py propagates over that entity graph directly (Collins & Loftus 1975). Memory tags exist as flat classification labels on a memory row, not as an addressable intermediate node type bridging cues to contents.

Why the layer matters

Their construction is two LLM distillations per episodic unit e_i:

• g_i = F_LLM^tag(e_i) — an associative tag summarising the relational pattern of the episode
• C_i = F_LLM^cue(e_i) — the fine-grained cues (entities, attributes)

Each cue links to the episode through the tag. Retrieval is then two-stage: select relevant tags, then fetch content conditioned on the selected tags. The paper is explicit that this is what prevents "combinatorial explosion caused by unconstrained expansion" — the failure mode their Figure 2 shows for plain graph expansion, which "retrieves additional but irrelevant neighbors, yet still fails to recover information about Caroline."

This is also where their token win comes from: "MRAgent maintains a lightweight construction phase and defers complex relation-building to the retrieval stage, where it is performed in a query-specific manner." Cortex does the opposite — heavy write-time structure building in core/curation.py and core/knowledge_graph.py.

Their justification is engram-theoretic (Frankland & Josselyn 2019; Rashid et al. 2016), with tags as "intermediate associative structures" analogous to engram reactivation — the same literature core/engram.py already cites via Josselyn & Tonegawa (2020). So this is an extension of a lineage Cortex is already in, not a foreign graft.

Ask

Introduce an addressable associative-tag node type between cues and contents, with the three traversal operators (Cue→Tag, (Cue,Tag)→Content, Content→(Cue,Tag)) as the graph API #361 consumes.

Constraints

• This is a schema change to a persisted store: one-shot migration, no back-compat shim (standing lead rule, §13 E3). Old-data-in / new-code-out must have a test.
• Both backends must be covered. A green SQLite suite does not establish PostgreSQL behaviour (§12.3 — the numpy.float32/structuredContent incident).
• Land as its own PR before recall: the read path is static retrieve-then-rerank — no evidence-conditioned reconstruction loop (MRAgent, +23% LoCoMo) #361 (§15: prerequisites are built and sequenced, never logged).

Acceptance criteria

• Tag nodes are addressable and the three operators are implemented and unit-tested on both SQLite and PostgreSQL.
• A migration converts an existing store with row counts asserted before and after — not merely "no exception raised" (silent-success trap).
• Write-path cost measured: the two extra LLM distillations per episode are a real ingest cost, so the before/after ingest latency is recorded and either accepted in writing or offset.
• Zero surviving non-equivalent mutants on changed files (§12).