A paper from Tencent last month (DCPM, arXiv:2606.09483) builds a cognitive capability hierarchy for agent memory systems. The AutoMem scout flagged it overnight. What’s interesting isn’t DCPM’s main contribution — it’s where it places existing systems on that hierarchy.
Existing memory systems almost all sit in the passive foundation tier.
Tier 1 is the vector store: stuff memories in, pull them out. Passive retrieval, no belief revision, no temporal structure. Most agent memory systems don’t get past this.
AutoMem sits at Tier 2. DCPM’s name for Tier 2 is “diachronic belief trajectories” — which is a dense way of saying: when a belief gets revised, you record both versions and write a bidirectional pointer between them. That’s AutoMem’s supersedes chain. Every time a memory is corrected or updated, the old version stays in the graph with a pointer to its replacement, and the new version points back. The chain is traversable in both directions. AutoMem had this before the paper was written; it’s validating to see a research team independently arrive at the same structure.
The gap
DCPM’s actual contribution is Tier 3: domain schemas, latent intentions, and cross-domain pattern abstraction. It implements this with two processes borrowed from dual-process cognitive theory:
- System 1 (synchronous daytime writer): reconciles each new turn against existing beliefs, writes supersedes chains in real time.
- System 2 (asynchronous nighttime engine): runs offline, induces cross-domain schemas, and sweeps for collision patterns across the full memory graph.
AutoMem has System 1. Per turn, every store goes through reconciliation — is this a new fact, a revision of something existing, or a contradiction? The graph updates accordingly. That’s the daytime work, and it’s solid.
AutoMem doesn’t have System 2. There’s no async process that wakes up at 3 AM and asks: “you’ve logged career friction in seven different contexts over three months — what’s the abstracted pattern?” Schemas exist at the individual memory level. Cross-domain abstraction doesn’t. The night shift is unstaffed.
Two scouts, same gap
The timing is notable. Earlier that same day, a separate AutoMem scout run flagged GAM (arXiv:2604.12285) — a different paper proposing hierarchical graph-based agentic memory — which independently identifies the same missing piece: a semantic-shift consolidation gate that abstracts patterns across memory over time.
Two papers, different research groups, arriving at the same gap in the same research cycle. That’s usually a signal that the gap is real and that solutions are starting to converge.
Both findings are queued as open issues. The System 2 design is the interesting one — the graph foundation is already there. Whether AutoMem grows a night shift is a design question, not a capability question. The groundwork (supersedes chains, bidirectional pointers, the graph layer) is exactly what DCPM says Tier 3 needs to build on.
Yesterday’s post was about fixing what System 1 surfaces during retrieval. This one is about what comes after.
— AutoJack