Episodic Memory and Event Histories

Some questions have answers that expire. FreshQA is a 600-question benchmark built around exactly that. Its questions are split into never-changing, slow-changing, fast-changing and false-premise. The fast-changing slice is where a once-correct answer keeps being returned after it has stopped being correct.

Models do badly there. Across the models tested, overall accuracy ran from 0.8% to 32.0% under the strict grading that penalises hallucination, and from 0.8% to 46.4% under relaxed grading. On the fast-changing questions the strongest model collapses. The 2024 FreshLLMs paper says so on its own results: “While GPT-4 generally obtains the highest accuracy on these questions, it never exceeds 15% in both evaluation modes.”

The refusal behaviour is the other half of the picture. On those same fast-changing questions GPT-4 declined to answer 60% of the time, against ChatGPT's 16%. Read the two numbers together and the design problem is exact. A system that cannot tell whether a stored fact is still in force must choose between answering wrongly and refusing broadly. Neither is a memory policy. A record that carried its own time, policy and environment conditions would not have to choose.

Episodic memory represents particular events: a task, observation, action, result, approval, failure, or human correction. It can help agents recognize recurring situations and avoid investigating the same thing twice. Reflexion is a working instance of that contract, and it keeps the episodes outside the weights. Its 2023 abstract says how: “Reflexion agents verbally reflect on task feedback signals, then maintain their own reflective text in an episodic memory buffer to induce better decision-making in subsequent trials.”

The headline figure is 91.0% pass@1 on HumanEval Python, against 80.1% for GPT-4 and 65.8% for the prior state of the art (CodeT + GPT-3.5). The number that matters for record design is smaller and sits in an ablation. On 100 HotPotQA questions the authors separated self-reflection from plain episodic memory — the condition that simply replays the most recent trajectory. Self-reflection was worth an 8% absolute boost over replay alone. Storing the run is not the same as storing what the run means. The gap between them is measurable.

Past success is not a universal rule. Episodes need time, context, version, scope, and outcome so retrieval can distinguish precedent from obsolete accident.

Healthy retrieval shows itself in four ways, and a public benchmark now puts numbers on how hard each one is. LoCoMo, from Maharana and colleagues, is a dataset of 10 very long-term dialogues, averaging 588.2 turns, 27.2 sessions and 16,618.1 tokens each, running up to 32 sessions. Its authors open by saying what was being measured before: “Existing works on long-term open-domain dialogues focus on evaluating model responses within contexts spanning no more than five chat sessions.” Five sessions was the old horizon. LoCoMo runs to 32.

The QA benchmark on top of it carries 1,986 questions. Of those, 321 (16.1%) require temporal reasoning and 446 (22.4%) are adversarial. Long-context LLMs and RAG improved QA by 12-20% and still lagged human levels by 36% overall, and by 41% on temporal reasoning. That 41% falls on precisely the judgement that decides whether a retrieved episode is still current. On the adversarial questions long-context models scored 65% lower than the base model. That is what Signal 3 looks like when it fails: more history retrieved, worse discrimination about which of it applies.

An episode worth storing has four parts — Episode identity, Event sequence, Outcome, and Reflection — plus the retrieval metadata that decides when it comes back. Only Reflection is an interpretation. The cleanest demonstration that these are genuinely separable is a system that separates them, then removes one to see what breaks.

That system is 25 generative agents living in a sandbox town, built by Park and colleagues in 2023, with episodic memory as a memory stream of timestamped memory objects. Retrieval is a weighted sum of three terms. Recency is an exponential decay with factor 0.995. Importance is an LLM-assigned 1-10 poignancy score. Relevance is cosine similarity to a query memory. All three weights are set to 1. That is the retrieval-metadata layer written out as arithmetic.

Reflection sits above the raw observations without replacing them: “Reflections are higher-level, more abstract thoughts generated by the agent. Because they are a type of memory, they are included alongside other observations when retrieval occurs.”

The ablation then prices each part. Rated for believability by TrueSkill, the full architecture came first at mu = 29.89, sigma = 0.72. Removing reflection dropped it to 26.88. Removing reflection and planning, to 25.64. The human crowdworker condition scored 22.95, and the baseline with no memory, planning or reflection scored 21.21 — a standardised effect size of d = 8.16 between that baseline and the full architecture (Kruskal-Wallis H(4) = 150.29, p < 0.001). Give Reflection and Retrieval metadata separate owners: the ablation prices Reflection on its own, while the retrieval weights were never ablated, so only a test of their own will show what they contribute.

Episodic records are easiest to design for a workflow whose past incidents you can still read. The record earns its place when it makes an old run reusable without making it binding. Call up a case that worked once, and the record should still say when it happened, under which policy, and in what environment. That is what a reader needs to judge whether the outcome carries over to the task in hand.

A support agent reused a workaround from an incident that had already been fixed. The memory search surfaced a successful past episode with similar wording, and omitted that the workaround applied only during a temporary outage and was prohibited after the service recovered. Three kinds of stored experience are compared. A raw event log is the immutable, time-stamped record of what was observed and done. An episode summary is a later reading of one episode, linked to those events. A behavioral precedent is a past episode retrieved as a template for what to do now. The incident, times and policy versions below are invented for the exercise.

Keep raw events immutable and interpretations revisable. This separation supports audit, improved reflection, and correction without rewriting history. In regulated systems it is not a preference. It is codified. The FDA wrote it into its electronic-records rule in 1997, and 21 CFR 11.10(e) states it in one paragraph: “Use of secure, computer-generated, time-stamped audit trails to independently record the date and time of operator entries and actions that create, modify, or delete electronic records. Record changes shall not obscure previously recorded information. Such audit trail documentation shall be retained for a period at least as long as that required for the subject electronic records and shall be available for agency review and copying.”

Read that as a specification for an episodic store and it gives you three clauses. The audit trail is computer-generated and time-stamped, so provenance is not something the agent narrates about itself. A change may not obscure what was previously recorded, so a revised reading is appended, never substituted. And the trail is retained at least as long as the record it describes — the clause that keeps an old episode auditable for exactly as long as it remains retrievable.

The habit worth carrying into review is caution about precedent. A retrieved success can be the wrong precedent, because the run behind it took place under conditions that may since have changed. So before an old run is reused, check two things. That its record still carries those conditions. And that they match the situation in front of you.