MCP consolidate_now can never take the FlashAttention path — learner constructed without config (agentdb-mcp-server.ts:282)

Summary

consolidate_now over the MCP server can never take the FlashAttention consolidation path, so it reports episodesProcessed: 0 on every call and discovers no causal edges — regardless of how much episode data exists.

I understand ENABLE_FLASH_CONSOLIDATION is a deliberate feature flag defaulting to false (NightlyLearner.ts:17, :78). The issue isn't the default — it's that the MCP server provides no way to opt in, and the fallback is silent, so the tool looks like it succeeded.

Root cause

src/mcp/agentdb-mcp-server.ts:282 constructs the learner with no config argument:

const learner = new NightlyLearner(db, embeddingService);

It therefore inherits the constructor default (NightlyLearner.ts:78) where ENABLE_FLASH_CONSOLIDATION: false. The AttentionService is only created when that flag is set (:88-89), so attentionService stays undefined and consolidateEpisodes() takes its fallback branch on every call (:213), returning episodesProcessed: 0.

The consequence is that the reward-delta discovery loop — the only code that converts episode outcomes into causal edges — is unreachable through the MCP surface.

Evidence

Environment: agentdb@3.0.0-alpha.20 (resolved via agentic-flow), MCP server started with agentdb mcp start, store with 2993 episodes across 2 sessions.

Before — every consolidate_now call, session-scoped or global:

{ "edgesDiscovered": 0, "episodesProcessed": 0, "experimentsCreated": 10,
  "recommendations": ["No new causal edges discovered. Consider collecting more diverse episode data.", ...] }

SELECT COUNT(*) FROM causal_edges → 0.

The recommendation points at the data, which sent me looking in the wrong place for a while — the data was never read.

After — same database, same episodes, learner constructed identically except for the flag:

new NightlyLearner(db, embedder, { ...defaults, ENABLE_FLASH_CONSOLIDATION: true })

attentionService present: true
episodesProcessed: 1000
causal_edges: 0 -> 94

Verified by direct SQL rather than the tool's own report:

check	result
mechanism	flash_attention_consolidation on all 94
uplift	signed, −0.35 … +0.35
|uplift| >= 0.05	94 / 94
PRAGMA integrity_check	ok

Suggested fix

Pass a config at the construction site, ideally with an env opt-in so operators can enable it without a code change:

const learner = new NightlyLearner(db, embeddingService, {
  ENABLE_FLASH_CONSOLIDATION: process.env.AGENTDB_FLASH_CONSOLIDATION === '1',
});

Two smaller observations, take or leave

The fallback is silent. episodesProcessed: 0 is indistinguishable from "there was nothing to process". Since the fallback is a real capability difference, it may be worth surfacing it in the report — e.g. mode: "fallback" or a note that flash consolidation is disabled. That one line would have saved me most of the investigation.

A downstream call site overrides the threshold. agentic-flow's dist/reasoningbank/AdvancedMemory.js:39 constructs NightlyLearner with upliftThreshold: 0.1, double the library default of 0.05 (NightlyLearner.ts config default). In my corpus, 12 of the 94 edges clear 0.05 but not 0.1. Not an agentdb bug — flagging it in case the divergence is unintentional.

Thanks for agentdb — the causal-edge model is genuinely useful, and once the path actually runs it produces exactly the signed-uplift structure the docs describe. Happy to open a PR for the one-line change if that's helpful.

Start at the learner construction in src/mcp/agentdb-mcp-server.ts:282, then read the defaults and AttentionService setup in NightlyLearner.ts:17,78,88-89 and the fallback at :213. Exercise the MCP consolidate_now entry point with and without the relevant environment setting. Done means operators can opt into flash consolidation and the result reports processed episodes when episode data is available.