Tencent Team Open-Sources 'T-Mem' Memory System with Rehearsal Mechanism to Fill Associative Recall Gap

T-Mem's problem awareness stems from a unified judgment about existing architectures: Flat RAG, graph-structured memory (GraphRAG, Zep, HyperMem), agentic hierarchical memory (Mem0, A-Mem, MemoryBank), and operating-system-style memory kernels (MemGPT, MemOS, MIRIX)—though structurally different—share an identical retrieval recipe: project the query and stored content into the same similarity space (BM25 or dense vectors) and take top-K. According to the release content, this determines one thing: memory "reachability" is capped by similarity.

The paper breaks this ceiling into two types of recall. When the query and memory share surface features (the same wording, the same entity, the same time and place), similarity retrieval works well—this is descriptive recall. But long conversations contain an equally common other half: the query and memory share zero surface overlap, bound only by a latent semantic arc, such as a causal relationship or the continuation of the same situation—this is associative recall. The example given in the release content: a month ago the user said "someone on the team has a severe seafood allergy," and today asks "where should the team eat tonight"—the two sentences share no vocabulary overlap, yet the former is precisely the necessary basis for answering the latter.

This yields a 2×2 retrieval design space: "granularity" (single fact / complete conversation segment) and "orientation" (descriptive / associative) as two orthogonal axes. The paper points out that existing systems all crowd into the first and fourth quadrants—the descriptive half; the associative half, the second and third quadrants, is the structural blind spot of this retrieval recipe. This is T-Mem's design starting point.

T-Mem's answer borrows from "episodic future thinking" in cognitive science—when people recall the past, they rehearse experiences for future cues. Its engineering counterpart is the trigger: when a memory is written, a prediction computed offline by the builder LLM and stored alongside the host, answering "under what circumstances this memory will become important again." Its architectural significance lies in decoupling two things: "how a memory is reached" and "what counts as answer evidence." When similarity cannot find the host, the query can still hit it via one of the host's triggers, so each memory retains both a descriptive entry point and an associative entry point.

The four trigger families each occupy one quadrant of the design space. Entity Trigger (fact × descriptive) gives an atomic fact a superordinate concept name—for example, the entity trigger for "seafood allergy" might land on "dietary restrictions"; Bridge Trigger (fact × associative) projects a fact onto a specific situation where "knowing it would be useful" and attaches a one-step reasoning rationale—for example, projecting the allergy fact onto "choosing a restaurant for the team dinner"—this is the pure associative axis, completely unreachable by similarity retrieval; Scene Trigger (scene × descriptive) describes the current scene with one sentence each from four orthogonal attributes: situation, object, event, and emotion; Horizon Trigger (scene × associative) projects the same scene onto a set of forward-looking dimensions, so that future queries approaching from related but different situations can still hit it. Among these, Entity and Scene serve the half already covered by similarity retrieval; Bridge and Horizon are the two families that fill the blind spot.

Memories are organized into a typed graph of five object classes: scene (coherent conversation segments split by event closure, serving as evidence), item (atomic facts extracted from scenes, anchorable to multiple source scenes to carry cross-scene logical chains), topic tags (used only to delimit extraction scope and retrieval pre-filtering, explicitly not entering the QA channel), four trigger families (participating only in retrieval, never appearing in evidence), and persona aggregated by speaker (attached after evidence as environmental context, not consuming retrieval budget). The three design commitments are: evidence layer isolated by type, topics do not pollute QA, and triggers do not enter the evidence path.