Dedup Is Not Epistemics

All numbers in this post are machine-generated and committed: demo + evidence.*

Agent memory vendors sell consolidation as understanding, but it isn't. I planted 20 facts in a LangMem store and ran their consolidation with their consolidation directives verbatim. It merged 4 of 5 duplicates and caught 0 of 14 contradictions. Given the chance to see the conflicts, it deleted the evidence instead. This post is about my investigation.

Disclaimers up front: this tests langmem 0.0.30, which is pre-1.0 and says so. The langgraph stack around it is post-1.0 and was rock solid. The critique is about what consolidation IS, not about ship quality. And daftari, the other tool in this post, is my project; discount accordingly, then run the three commands and check.

The tell: pricing memory as inventory

Look at how agent memory gets priced. ContextVault's tiers are gated by memory count: 50 memories on the trial, 500 at $9.99 a month, 2,500 at $49.99, unlimited on Enterprise. You upgrade when you accumulate. Their marketing says nothing about deduplication, reconciliation, or consistency. This isn't about one vendor; it's the category's business model showing. A warehouse charges for pallets. It doesn't audit what's on them, and it has no reason to tell you two pallets contradict each other.

Inventory pricing is at least understandable. The problem starts when a vendor prices like a warehouse and markets like an auditor.

The claim under test

LangMem's default prompt instructs its memory manager to "remove incorrect or redundant memories while maintaining internal consistency." That's an epistemics claim. Consistency maintenance, in writing, in the default prompt.

So I tested it as written. One honest disclosure: their default prompt extracts user-profile memories, so I used their documented instructions parameter to point extraction at org facts instead. The consolidation directives, the part under test, stayed verbatim. Every prompt is in the repo.

I ran four simulated agent sessions for a fictional company: pricing, ops, support, docs. I planted 20 facts among 30 benign ones. Five near-duplicate pairs as the control group. Three pairwise contradictions. Two temporal traps. And one four-way contradiction where no single pair reveals the full inconsistency:

• Pricing sells 500 requests/sec, guaranteed on order forms.
• Ops hard-caps the gateway at 350.
• Support grants 800 rps bursts without asking ops.
• Public docs say every plan throttles at 200.

Embedding distance between these claims: 0.48 to 0.66 cosine. The near-duplicate pairs sit at 0.81 to 0.93. Vector similarity will never pair the contradictions. You need structure, not similarity.

What their consolidation did

The control worked: 4 of 5 near-dup pairs merged. Their dedup is real.

The contradictions: 0 of 14 caught, in both full runs, across three configurations (their documented per-agent namespaces, a shared namespace, and a global single-context pass). Two failure modes, and you always get exactly one of them:

1. Namespace-scoped (their documented pattern): blind. Consolidation can't see across agent namespaces. All 14 contradictions survive, unflagged, forever.
2. Shared namespace (the charitable setup): destructive. It saw the conflicts and resolved them by deleting one side. The 500 rps revenue guarantee: gone. The us-east-1 deploy region: gone. Recency won every arbitration. No flag, no tombstone, no audit trail. It also rewrote two conflicting capacity claims into one harmonized policy that nobody in any session ever stated.

Bonus finding: extraction invented a fact. It promoted the fictional company's CTO to CTO of their biggest customer, and consolidation kept the fabrication. Your memory layer generates confident wrong facts, then defends them against correction.

The counterexample that matters

Daftari read the same Postgres store through a read-only role and compiled each memory into a claim note with provenance back to the store row. Then an agent pass judged related notes for conflict. Same model as LangMem's run. Same one-pair-at-a-time retrieval scope. No global view, ever.

Result: 3 of 3 pairwise contradictions logged, both temporal traps flagged without being auto-resolved by recency, and at most one borderline flag on the 30 benign facts across two runs (a filler fact that turned out to genuinely tension with a planted one; it's in the committed report, since I believe in transparency).

And the four-way capacity conflict surfaced as one connected component spanning all four sessions, assembled from three pairwise judgments that never saw each other. The graph knows something no single LLM call knew.

That's the argument in one sentence: the intelligence isn't in the judge, it's in the ledger. LangMem had the same judge and destroyed the evidence. An append-only tension log turned three narrow observations into an organizational diagnosis: your pricing, ops, support, and docs teams are selling four different products.

The retrofit play

Nobody migrates memory stores. You don't have to. LangMem's store is a Postgres table; reading it is the whole integration. Daftari imported the memories without LangMem's runtime, without write access, without touching the producing system. One command, 49 claim notes, every note traceable to its source row. Any memory store with a readable substrate gets the same treatment: compile the claims, keep the provenance, grow the tension graph on top.

Dedup makes stores smaller but epistemics makes them honest. Different products. The second one retrofits onto the first.

If your memory vendor's consolidation "maintains internal consistency," ask them one question. When two memories conflict, where does the loser go? There are only three answers. It survives untouched (you have a search index, not memory). It disappears (you have evidence destruction). Or it's recorded as a conflict (you have epistemics).

Price accordingly.

---

Reproduce it: three commands from a clean clone. Full planted-vs-detected tables: RESULTS.md. Think I misconfigured LangMem? The fixtures are committed and the repo takes PRs. Show a configuration that catches the plants and I'll publish the correction with your name on it.