Long-running agents must retrieve relevant information from histories in which facts change over time. Flat-context prompting is one approach, while structured temporal memory and selective retrieval offer alternatives. This proposal tests whether a learned event selector can improve temporal question answering under a fixed context budget, beyond strong retrieval and deterministic baselines.
The motivation is IBM Technology's Context Engineering for AI Agents. It names four ways context fails long-running agents — poisoning, distraction, confusion, clash. This idea tests only the temporal slice: retrieval, revision, and answer accuracy.
The research question
Does a learned selector, constrained by explicit temporal rules, improve the accuracy–cost trade-off over strong retrieval and a well-designed deterministic query, under equal context budgets?
The pipeline: an append-only event log (type, time, entity, payload, and revision links keyed by fact or attribute — not by entity alone) feeds deterministic filters that narrow the log to candidates; a fine-tuned decision model ranks the events needed to answer; a frozen Qwen3-0.6B answers from the selection. The learned selector is the component under evaluation — the event store and filter are infrastructure, and the decision model is fine-tuned, not Qwen.
APEX-MEM (temporally grounded event graphs, retrieval-time resolution), Chronos (structured temporal events, query-guided retrieval), and ARTEM (spatio-temporal episodic memory and retrieval) already cover much of event-based memory and temporal retrieval, and learned selection exists in sparse attention and differentiable retrieval. The contribution here is a controlled comparison of learned evidence selection against strong deterministic and retrieval baselines under a fixed context budget — not the introduction of event-based memory.
The experiment
I will evaluate synthetic temporal question answering over event histories containing facts, revisions, and retractions.
Data. Split by event-history template or scenario, not random questions; hold out combinations of revision, retraction, temporal query, and unrelated updates, so the selector can't memorize patterns.
Method. Question plus typed historical events (timestamps, revision links) go in. Deterministic temporal filters remove ineligible candidates. A fine-tuned decision model — trained to identify the evidence useful for each question — ranks the remaining events within a fixed token budget, and a frozen Qwen3-0.6B generates the answer. The decision model is fine-tuned; Qwen stays frozen. Queries distinguish "what is true now" from "what was true at t" so recency alone can't win; for the first experiment, order events by a monotonic sequence number with explicit query-time semantics, deferring occurrence-vs-recording.
Selector model. A pretrained decision model is the starting point; Tev1 is a candidate, not a confirmed backbone. Its standard interface selects a single option, so the initial experiment tests single-event selection; multi-event selection under a token budget follows only if that succeeds. (Its LoRA recipe targets Qwen3.5-4B; verify the separate 0.8B variant is supported.)
Baselines. Full context, embedding retrieval, recency selection, a deterministic revision-aware selector, and the original (un-fine-tuned) Tev1 through the same selection interface — under comparable token budgets. The deterministic selector resolves revisions per fact or attribute, respects the question's temporal reference, and keeps the evidence needed to interpret a correction or retraction. The Tev1 control separates what fine-tuning adds from what a pretrained decision model already provides.
Evaluation. Fix an equal context budget — the same maximum generator-input tokens, same tokenizer, counting question, instructions, and selected events. Measure answer accuracy as history length increases, evidence-selection precision and recall, actual tokens, and end-to-end latency including selector inference, retrieval, and serialization. First verify that the generator can answer correctly when given gold evidence.
Falsification. The hypothesis is unsupported if the learned selector does not improve the accuracy–cost trade-off over strong baselines. If a deterministic method performs equally well at lower cost, prefer the simpler method.
Limitations
Rules may suffice; selection can drop the statement a correction fixes; overhead can erase the token savings; and generalization beyond synthetic data is open. Fine-tuning the generator Qwen, end-to-end gradients, and learned operators are deferred until the frozen-generator run shows a gap.
References
Related work
- APEX-MEM — temporally grounded event graphs, append-only storage, retrieval-time resolution.
- Chronos — timestamped event tuples for time-sensitive QA.
- ARTEM (AAAI) — spatio-temporal episodic memory and retrieval.
Experimental backbone
- Qwen3-0.6B — Hugging Face.
- Tev1 (0.8B) — fine-tuned decision model; LoRA recipe (Qwen3.5-4B) in the Together AI repo.