Give users control over memory capacity and forgetting
A controllable memory isn't an infinite archive—it's one where you decide what stays and what fades. A 2024 model called DyMemR introduced a 'human-like memory pool' with explicit capacity limits, memory loss, and repetition stimulation [1]. This means the system automatically forgets less important facts and emphasizes ones that recur, which improved reasoning accuracy over baselines. For a user, this translates to controls like 'set a memory size' or 'boost important events'—so the graph doesn't drown in noise.
The same study found that using all historical facts introduces noise and invalid information, hurting reasoning [1]. So, a controllable design should let users filter or prune old data, not just accumulate everything. This is a direct design choice: expose a 'memory retention' slider or a 'forget old events' button, backed by evidence that selective memory beats exhaustive storage.
Let users retrieve and synthesize across time, not just recent events
Control also means being able to ask 'what happened last month?' and get a synthesized answer, not just a list of recent facts. The Zep system, a 2025 temporal knowledge graph architecture, excels at cross-session information synthesis and long-term context maintenance [2]. In the LongMemEval benchmark, it improved accuracy by up to 18.5% and cut response latency by 90% compared to baselines [2]. For a user, this means having a search box that understands time—like 'show me all meetings with X in Q3'—and returns a coherent summary, which feels far more controllable than a raw dump.
Another study, MTDM, found that focusing only on recent facts overlooks earlier ones that hold useful knowledge [3]. So, a controllable design should offer a 'time range' filter or a 'historical context' toggle, letting users pull from any period. This is a concrete design choice: make temporal filters explicit and easy to adjust, so the agent's memory doesn't feel like it has amnesia.
Offer control over how the graph reasons, not just what it stores
Controllability isn't only about storage—it's about how the agent uses the graph to answer. A 2023 model called DaeMon models temporal path information between query subjects and object candidates without depending on entity representations [4]. This means users can control reasoning by specifying paths or relations, not just entities, which is useful when new entities appear over time. For a user, this could be a 'reason by relation' mode, letting them ask 'what led to this event?' and get a path-based explanation.
This design choice is supported by the fact that DaeMon outperformed state-of-the-art models by up to 4.8% in MRR (mean reciprocal rank, a measure of how well the model ranks correct answers) [4]. So, giving users control over the reasoning mechanism—like choosing between entity-centric and path-centric views—can make the memory feel more transparent and steerable.
About These Sources
This answer is built on 4 studies (3 peer-reviewed, 1 preprint) — published from 2023 to 2025, 2 from 2024 or later, 1 in Q1 journals — selected as the most relevant from 4 studies that passed quality screening, drawn from 35 papers retrieved from a database of over 500 million.
Sources used in this answer
Temporal Knowledge Graph Reasoning With Dynamic Memory Enhancement
DyMemR, a 2024 TKG reasoning model, introduced a human-like memory pool with capacity limits, memory loss, and repetition stimulation, and showed that selective memory (vs. using all history) improves reasoning accuracy.
Zep: A Temporal Knowledge Graph Architecture for Agent Memory
Zep, a 2025 temporal knowledge graph architecture, outperformed MemGPT on the Deep Memory Retrieval benchmark (94.8% vs 93.4%) and improved accuracy by up to 18.5% while reducing latency by 90% on the LongMemEval benchmark, particularly for cross-session synthesis and long-term context.
Temporal knowledge graph reasoning triggered by memories
MTDM, a 2023 memory-triggered decision-making network, showed that using earlier historical facts in addition to recent ones improves prediction accuracy and efficiency, and that focusing on the most relevant facts mitigates imprecise knowledge propagation.
Adaptive Path-Memory Network for Temporal Knowledge Graph Reasoning
DaeMon, a 2023 adaptive path-memory network, models temporal path information without depending on entity representations, achieving up to 4.8% absolute improvement in MRR over state-of-the-art models on four real-world TKG datasets.