1 Introduction↩︎

A persistent limitation of large language model (LLM) agents is the absence of a continuously evolving memory. The context window provides short-term recall within a session, but once a session ends all conversational history is lost unless externally persisted. This has motivated a growing body of work on agent memory systems, mechanisms that store, retrieve, and reason over information accumulated across many sessions.

Existing approaches treat memory as a database of facts. Retrieval-augmented generation (RAG) [4] encodes conversation chunks as dense vectors and retrieves the \(k\)-nearest at query time. Knowledge-graph systems such as Zep [5] store subject-predicate-object triples. Episodic systems such as MemGPT [6] maintain explicit core memory strings and flush older records to an archival store. More recent systems including Mem0 [7] and A-MEM [2] combine graph and vector representations to improve retrieval precision.

Despite their diversity, all these systems share one assumption: memory is a collection of things that happened. This produces three structural failure modes. First, conflict blindness: when a user changes jobs or preferences, contradictory records accumulate with no principled resolution mechanism. Second, heuristic forgetting: deletion rules are time-based rather than semantic. Third, retrieval granularity mismatch: a question asking what a person has painted over the years requires aggregating values from many sessions, but a fact store returns individual records with no accumulation mechanism.

We argue these failures share a root cause: treating memory as storage rather than as a model of the world. The brain does not store a verbatim transcript of past conversations; it maintains a generative model of its environment and updates it when observations diverge from predictions [8], [9]. Surprise drives learning. Unused beliefs decay toward uncertainty. Memory is a compressed, continuously refined probability distribution over the state of the world.

Nous (from the Greek word for mind or intellect) implements this view as a practical agent memory system. The core data structure is a dimension, a categorical distribution over the possible values of a single entity-attribute pair. New observations are scored by information-theoretic surprise and integrated via Bayesian posterior update. The record of that update, a delta, is the primary stored artifact, capturing not what was observed but how it changed the agent’s understanding. Forgetting is entropy decay, and conflict resolution is implicit: contradictory evidence naturally shifts probability mass away from the old value.

Our contributions are: (1) the predictive world model paradigm for agent memory, formalised via dimensions, deltas, and surprise-driven Bayesian updates; (2) a complete open-source implementation requiring no external database; and (3) an evaluation on LoCoMo demonstrating strong absolute performance, presented alongside an explicit, honest discussion of the inconsistencies we found across the literature when attempting to source comparable baseline numbers, including a head-to-head comparison against BeliefMem [3], a concurrent system built on a closely related premise.

2 Background and Related Work↩︎

2.0.0.1 Long-term conversational memory.

The LoCoMo benchmark [1] evaluates agent memory over conversations spanning up to 35 sessions, 300 to 600 turns, and 9,000 to 16,000 tokens, covering single-hop, multi-hop, temporal, and open-domain questions. LongMemEval [10] (ICLR 2025) extends this with harder, audited questions and is increasingly adopted as a complementary standard.

2.0.0.2 RAG-based memory.

RAG [4] retrieves the top-\(k\) embedding-nearest chunks. Its primary failure is the inability to aggregate information distributed across many chunks, compounded by the lost-in-the-middle effect [11] wherein relevant evidence buried in long contexts is ignored.

2.0.0.3 MemGPT.

MemGPT [6] partitions memory into in-context, recall, and archival tiers managed via function calls. It is effective for single-session recall but lacks conflict-resolution semantics.

2.0.0.4 A-MEM.

A-MEM [2] enriches memories with Zettelkasten-style links, forming a graph of interconnected notes with LLM-generated attributes. It is a widely-cited academic baseline on LoCoMo. We discuss its reported numbers, and a substantive discrepancy we found across the literature, in Section 5.

2.0.0.5 Zep and Mem0.

Zep [5] builds a temporal knowledge graph; Mem0 [7] maintains a hybrid vector and graph store. Both report strong LoCoMo scores in their own evaluations but typically use GPT-4o with an LLM-as-judge scoring protocol rather than strict token F1. A community audit found this judge protocol accepts semantically incorrect answers under adversarial conditions a non-trivial fraction of the time, and that a small percentage of LoCoMo ground-truth labels themselves contain errors [12]. Because Nous is evaluated with strict token-F1 under a different backbone, we do not present a head-to-head numeric comparison against Zep or Mem0 in this paper; we report their published numbers only as orientation for the reader in Section 5.

2.0.0.6 Belief-based memory.

Concurrently with this work, BeliefMem [3] proposes storing, for each candidate fact, an explicit probability rather than a single deterministic value, arguing that deterministic memory collapses genuine uncertainty to a point estimate and produces what they term self-reinforcing error when that estimate is wrong: the agent acts on the stored conclusion, generates further observations consistent with it, and the error compounds. This is close to the conflict-blindness failure mode we identify above, reached independently and framed somewhat differently (BeliefMem frames it through a POMDP belief-state lens; we frame it through predictive coding and the free-energy principle). The two systems differ in three concrete design choices. First, BeliefMem deliberately maintains independent, unnormalised probabilities per candidate conclusion, explicitly rejecting a joint normalised posterior because it is intractable over an open-ended, dynamically expanding hypothesis space; Nous instead commits to a single normalised categorical distribution per entity-attribute pair (Definition 1). This makes Nous’s representation more tractable to reason about formally, but it is unimodal by construction, a limitation we discuss in Section 3.6 and Section 7 and one that BeliefMem’s design choice specifically avoids. Second, BeliefMem merges new evidence via a noisy-OR combination rule over an LLM-extracted, uncalibrated confidence score, while Nous uses a closed-form Bayesian posterior (Eq. 3 ) driven by an explicit, reportable surprise quantity in bits (Eq. 1 ); surprise is a first-class, logged signal in Nous rather than an implicit byproduct of the merge rule. Third, BeliefMem retains archived belief snapshots after each merge and scores retrieval by a similarity- and time-decay-weighted score over those snapshots, while Nous stores the belief revision itself (the delta, Definition 2) as the primary artifact, making every update individually auditable. BeliefMem is evaluated on both LoCoMo and ALFWorld, under both GPT-4o-mini and GPT-4o, with full ablations over its belief representation, retrieval mechanism, and update operations; this is a substantially more thorough empirical validation than the present preprint offers, and we discuss a direct LoCoMo/GPT-4o-mini comparison against their self-reported numbers in Section 5. Related but distinct, Belief Engine [13] applies a Bayesian belief-update framing to opinion dynamics in multi-agent debate rather than to conversational memory retrieval, and the broader POMDP belief-state literature [14] provides the theoretical grounding both BeliefMem and, via a different route, Nous draw on.

2.0.0.7 Predictive coding and the Bayesian brain.

Nous draws theoretically on predictive coding [8], which models the brain as a generative model minimising prediction error, and on the Free Energy Principle [9], which unifies perception and learning under surprise minimisation. The Bayesian brain hypothesis [15] grounds this computationally. Shannon’s information measure [16] supplies the definition \(\mathcal{S}(x) = -\log_2 P(x)\) bits, and Cover and Thomas [17] supply the entropy and mutual information foundations.

3 The Nous Architecture↩︎

3.1 Core Abstractions↩︎

A dimension encodes what the agent believes about one aspect of one entity. For instance, \(\mathcal{D}_{\text{Pranav},\text{employer}}\) might be \(\{(\text{Google}, 0.08), (\text{Sarvam AI}, 0.87), (\text{unknown}, 0.05)\}\). The mode \(\hat{v} = \arg\max_v p(v)\) is the current best belief.

The delta is the primary stored artifact of Nous. It records not what happened but how the agent’s understanding changed. The full delta sequence for a dimension is an auditable history of every belief revision.

3.2 Information-Theoretic Surprise↩︎

Given dimension \(\mathcal{D}_{e,a} = (V, p)\) and observed value obs, the surprise is:

\[\mathcal{S}(\text{obs} \mid \mathcal{D}_{e,a}) = -\log_2\, p(\text{obs}) \quad \text{bits.} \label{eq:surprise}\tag{1}\]

If obs \(\notin V\), the vocabulary is extended with \(p(\text{obs}) = \varepsilon\) and masses renormalised, yielding near-maximum surprise for novel values. An observation confirming the current belief causes negligible change; one that contradicts it forces a large revision, mirroring the role of prediction error in biological memory consolidation [9].

3.3 Bayesian Update↩︎

Nous computes the posterior via Bayes’ rule [18]. The likelihood is:

\[\mathcal{L}(\text{obs} \mid v) = \begin{cases} 1 - \varepsilon & v = \text{obs}, \\[4pt] \dfrac{\varepsilon}{\,|V|-1\,} & \text{otherwise,} \end{cases} \label{eq:likelihood}\tag{2}\]

and the posterior is:

\[p_{\mathrm{post}}(v) = \frac{\mathcal{L}(\text{obs} \mid v)\; p_{\mathrm{prior}}(v)}{\displaystyle\sum_{v' \in V} \mathcal{L}(\text{obs} \mid v')\; p_{\mathrm{prior}}(v')}. \label{eq:bayes}\tag{3}\]

This update is closed-form, runs in \(O(|V|)\) time, and requires no gradient computation. Multiple contradictory observations shift probability mass naturally without explicit conflict detection.

3.4 Entropy Decay↩︎

A dimension not updated for time \(\Delta t\) grows more uncertain via exponential mixing toward the uniform distribution:

\[p_t(v) = \lambda^{\Delta t}\; p_0(v) + \bigl(1 - \lambda^{\Delta t}\bigr)\; \mathcal{U}(v), \label{eq:decay}\tag{4}\]

where \(\mathcal{U}(v) = 1/|V|\) and \(\lambda \in (0,1)\) is a per-dimension retention factor. As \(\Delta t \to \infty\), \(p_t \to \mathcal{U}\), encoding complete uncertainty. The entropy satisfies:

\[H(p_t) \;=\; -\sum_{v} p_t(v) \log_2 p_t(v) \;\geq\; \lambda^{\Delta t}\, H(p_0), \label{eq:entropy}\tag{5}\]

guaranteeing the agent grows less certain as time passes without reinforcement, a property heuristic deletion rules cannot provide.

3.5 Identity Resolution↩︎

When two entity mentions may refer to the same person, Nous computes the symmetrised KL divergence between their shared dimensions:

\[\begin{align} D(e_1, e_2) = \!\!\sum_{a \in \mathcal{A}_{e_1} \cap \mathcal{A}_{e_2}} \!\!\!\!&\mathrm{KL}(p_{e_1,a} \,\|\, p_{e_2,a}) \\[2pt] {}+{} &\mathrm{KL}(p_{e_2,a} \,\|\, p_{e_1,a}). \end{align} \label{eq:kl}\tag{6}\]

Entities with low divergence across shared attributes are candidates for merging via a weighted mixture of their posteriors.

3.6 Observed-Values Aggregation↩︎

Bayesian posteriors are unimodal: after many updates, mass concentrates on one value, discarding historical alternatives. For inherently multi-valued attributes such as activities or places visited, this is incorrect. Nous addresses this via delta-log aggregation: at query time, every value that appeared with surprise above threshold \(\theta_{\min}\) is included in the context fact line:

\[V^*_{e,a} = \bigl\{ \Delta.\text{obs} \;:\; \Delta \in \mathcal{L}_{e,a},\; \Delta.\mathcal{S} > \theta_{\min} \bigr\}. \label{eq:obsvals}\tag{7}\]

This preserves the Bayesian world model for the current best belief while making the full historical record available for aggregation queries. We note this mechanism was added specifically to correct an early failure mode we observed during development, where multi-hop questions requiring an exhaustive list (e.g. "what activities does X do") returned only the single highest-probability value. We discuss this as an explicit limitation of pure unimodal categorical dimensions in Section 7 rather than treating it as fully resolved. We note that BeliefMem [3] addresses essentially the same underlying problem with a different and, in our view, more principled architectural choice: rather than aggregating post hoc from a delta log, it maintains independent per-candidate probabilities from the outset, so multi-valued attributes never require a separate aggregation step. Adopting an analogous representation is the highest-priority planned extension to Nous, discussed further in Section 7.

3.7 Pipelines↩︎

Figure 1 illustrates both pipelines. During ingestion, an LLM extractor parses each session into \((e,a,v)\) triples. For each triple, the current dimension is retrieved, the surprise computed via Eq. 1 , the posterior computed via Eq. 3 , and a delta appended to the log. A compressor prunes dimensions whose entropy exceeds a ceiling threshold.

During querying, named entities in the question seed a breadth-first search over the entity graph (up to two hops). For each retrieved entity, its dimensions are formatted as fact lines using Eq. 7 . The top-\(k\) delta records with the highest BM25 similarity to the question are appended as evidence lines, and the assembled context is passed to a category-specific prompt and then to the answering LLM.

4 Experimental Setup↩︎

4.1 Benchmark↩︎

We evaluate on LoCoMo [1]: ten dyadic conversations spanning 19 to 32 sessions each, totalling 1,540 questions across single-hop (841), multi-hop (282), temporal (321), and open-domain (96). The adversarial category is excluded following the canonical protocol used by prior work on this benchmark.

4.2 Metrics↩︎

We report token-level F1 and BLEU-1, both strict surface-form measures that do not reward semantically correct but differently-phrased answers. We additionally report context recall, the fraction of ground-truth answer tokens present anywhere in the assembled context, as a retrieval diagnostic, and a four-way failure bucket classification (good, answer miss, unknown, retrieval miss) assigned by a GPT-4o-mini judge.

4.3 Backbone and Baselines↩︎

All Nous experiments use GPT-4o-mini for extraction, answer generation, and judging. We compare against a no-memory baseline (direct QA with no retrieved context) and against A-MEM’s [2] self-reported GPT-4o-mini numbers. We do not run A-MEM ourselves; we rely on its published results and on independent citations of those results, and we report a discrepancy we found across the literature in Section 5 rather than resolving it unilaterally. We additionally compare against BeliefMem’s [3] self-reported GPT-4o-mini LoCoMo numbers under the same caveat: we do not run BeliefMem ourselves. We do not compare numerically against Zep or Mem0, as both are typically evaluated with a different backbone (GPT-4o) and a lenient LLM-as-judge protocol rather than strict token F1; their published numbers are reported in Section 5 for orientation only, not as a controlled comparison.

5 Results↩︎

Table 1 presents Nous’s results alongside a no-memory baseline and A-MEM’s self-reported numbers.

3pt

SH = single-hop, MH = multi-hop, T = temporal, OD = open-domain.

5.0.0.1 A caveat on the A-MEM comparison.

We sourced A-MEM’s GPT-4o-mini numbers from a table in a later paper that explicitly cites them as taken from A-MEM’s original publication [2]. We found that independent papers citing the same underlying A-MEM result disagree with each other on which numeric value belongs to which question category, while agreeing exactly on the four underlying numbers (27.02, 45.85, 12.14, 44.65). We report the category assignment we believe is correct based on cross-referencing two independent citing sources, but we flag this explicitly rather than presenting false precision: readers attempting to reproduce this comparison should consult A-MEM’s original paper directly rather than relying on any secondary table, including ours. Notably, BeliefMem [3] reports the identical four values for A-MEM’s official self-reported numbers (27.02 / 45.85 / 12.14 / 44.65 for multi-hop / temporal / open-domain / single-hop in their table ordering), which corroborates that these four numbers are the correct underlying values, even though it does not resolve which secondary source first assigned them to categories correctly. This same reproducibility gap appears to affect other papers in this literature and is, in our view, an under-discussed problem for the field.

Table 2 reports BLEU-1 and context recall for Nous in isolation, without the comparability concerns above.

2.5pt

The large BLEU-1/F1 gap on open-domain (10.06 vs 62.50) reflects that correct answers to personality and habit questions can be phrased in many ways, and BLEU-1 penalises valid paraphrases. On temporal questions BLEU-1 (56.04) and F1 (58.57) are nearly equal because temporal answers are precise dates or durations with minimal phrasing variation. The multi-hop gap (BLEU-1 23.90 vs F1 55.32) arises because Nous generates exhaustive lists covering all ground-truth items plus additional values, raising recall while lowering unigram precision, a direct consequence of the observed-values aggregation mechanism in Section 3.6.

5.0.0.2 Comparison against BeliefMem.

BeliefMem [3] is, to our knowledge, the closest existing system to Nous: both store per-attribute belief representations rather than deterministic facts, both update those beliefs from new evidence rather than overwriting them, and both report LoCoMo results under GPT-4o-mini against a shared baseline (A-MEM). This makes it a more informative comparison than the A-MEM comparison above. Table 3 reports both systems’ self-reported F1 scores on the same four LoCoMo categories, alongside the shared A-MEM baseline both papers report.

3pt

On these self-reported numbers, Nous exceeds BeliefMem on all four categories and on the macro average (59.97 vs.). We treat this as a genuinely informative comparison rather than a conclusive one, for several reasons we want to be explicit about rather than gloss over. First, "same backbone" does not mean "same pipeline": Nous’s extraction prompts, retrieval (BM25 plus two-hop entity BFS), and answering prompts differ from BeliefMem’s Add/Merge extraction and similarity-plus-decay retrieval over archived belief snapshots, and a gap of this size, especially the open-domain gap discussed below, plausibly reflects retrieval and prompting differences as much as the underlying belief-representation difference that is each paper’s central claim. Second, we have not verified that both papers evaluate strictly identical per-category question counts; we report \(n=841/282/321/96\) for SH/MH/T/OD in Table 2, and confirming BeliefMem uses the same counts is a check we have not yet completed. Third, and most importantly, BeliefMem’s own ablation (their Table 3) shows that removing their Merge operation, their analogue of Nous’s Bayesian update, drops their LoCoMo F1 from 42.38 to 20.38, and removing belief-based memory entirely drops it to 22.58. That is direct, controlled evidence, from a paper with ablations Nous currently lacks, that the shared underlying idea (revise beliefs from evidence rather than overwrite or discard them) is doing substantial work in their system. We read this as corroborating, not undermining, evidence for the broader direction Nous belongs to, independent of which specific instantiation ultimately performs better.

The open-domain gap is the result we are least confident in and the one we would most want a skeptical reader to interrogate. Nous’s 62.50 is more than double BeliefMem’s 28.73, yet, as discussed in Section 6, open-domain is simultaneously Nous’s highest-F1 category and its lowest-context-recall category (47.50%) and highest-retrieval-miss-rate category (37%) of any of the four. A high F1 alongside the weakest retrieval diagnostic in the entire evaluation is at minimum worth checking by hand against a sample of question-answer pairs before being reported as a confident win in any future revision, and we have not yet done this. We view a controlled, self-run comparison against BeliefMem under matched extraction and retrieval pipelines, ideally reported alongside the planned LongMemEval evaluation (Section 7), as the most important remaining experiment for determining which specific design choice, single normalised distribution versus independent per-candidate probabilities, Bayesian posterior versus noisy-OR merge, or delta log versus archived snapshots, is responsible for the difference observed here.

5.0.0.3 Orientation against industry systems.

For context only, and explicitly not as a controlled comparison, Zep and Mem0 report LLM-judge scores (not strict F1) under GPT-4o on LoCoMo in the range of roughly 60–80 across categories in their own published material. Direct comparison is not meaningful here given the different backbone, metric, and judge protocol; we mention this only so the reader can calibrate roughly where Nous’s strict-F1 numbers sit relative to that separate body of work, not to claim superiority over systems we have not run ourselves under matched conditions.

6 Analysis↩︎

6.1 Failure Bucket Breakdown↩︎

Table 4 shows the four-way failure distribution. Good denotes a correct answer; answer miss means context was present but the LLM answered incorrectly; unknown means the LLM declined to answer despite relevant context being present; retrieval miss means relevant information was absent from the assembled context.

3pt

For single-hop, the dominant failure is unknown answer (18%) despite a context recall of 86.07%. Of 154 such failures, the majority occur when context recall is well above the category average, meaning the information is present but the model declines to commit. This is consistent with a backbone calibration issue rather than an architectural retrieval failure.

For multi-hop, good (52%) and answer miss (33%) dominate the distribution at adequate context recall (75.35%), indicating the model still struggles in a third of cases to synthesise complete lists from distributed fact lines even when the underlying observed-values aggregation surfaces the right candidates. Retrieval miss (9%) contributes additionally, as some questions require linking entities beyond the two-hop search radius.

Temporal suffers from both the lowest context recall of all four categories (61.82%) and a high unknown-answer rate (19%). Temporal questions often depend on exact session dates embedded in evidence text rather than extracted as explicit dimension values, a representational gap we discuss in Section 7.

Open-domain has the highest retrieval miss rate (37%) and lowest context recall (47.50%) of any category. Questions about personality and long-term habits do not map cleanly onto entity-attribute dimensions, and improving retrieval for this category is the most promising remaining direction.

6.2 Context Recall as a Diagnostic↩︎

Context recall cleanly separates retrieval failures from answering failures. For single-hop, the 22.6-point gap between context recall (86.07%) and F1 (63.50%) identifies the answering step, not retrieval, as the binding constraint. For open-domain, context recall (47.50%) is itself close to the ceiling on F1 (62.50%), suggesting that architectural retrieval improvements for this category specifically would translate fairly directly into F1 gains.

7 Limitations and Future Work↩︎

We do not include controlled ablations isolating the contribution of individual components such as Bayesian update versus flat storage, entropy decay versus no forgetting, or delta aggregation versus current-state retrieval. Ablations are the primary planned extension; the bucket analysis above provides directional, not causal, evidence for each component’s contribution. BeliefMem [3] already provides this kind of ablation for its own, closely related, architecture (Section 5), and closing this gap is now the single highest-priority extension to this work.

We evaluate exclusively on LoCoMo in this preprint. LongMemEval [10] provides a harder, more carefully audited benchmark with fewer ground-truth errors, and evaluation on it is in progress and will be reported in a future revision. Community audits have identified erroneous ground-truth answers and judge-calibration issues in LoCoMo itself [12]; our use of strict token F1 rather than a lenient LLM-judge protocol is intended to be a conservative choice given these known issues, but it does not eliminate them.

As noted in Section 3.6, the underlying categorical dimension is unimodal by construction; the observed-values mechanism is an effective but somewhat ad hoc workaround for set-valued attributes rather than a first-class part of the probabilistic model. A dimension type maintaining an independent Bernoulli distribution per possible value, rather than a single categorical distribution over all values, is the highest-priority planned architectural extension and would let multi-valued attributes be modelled without a separate aggregation step; BeliefMem’s independent per-candidate probability representation (Section 5) is a closely related design already validated at scale and is a natural reference point for this extension. Making event timestamps explicit dimension attributes, rather than leaving them embedded only in evidence text, would additionally help close the temporal context-recall gap (currently the lowest of any category at 61.82%). All results in this preprint use GPT-4o-mini; confirming whether these gains generalise across other backbone models is necessary before drawing broader conclusions, and we plan to report this in a future revision alongside the LongMemEval results and the ablations described above. A controlled, self-run comparison against BeliefMem under matched extraction and retrieval pipelines, discussed in Section 5, is an additional and arguably more urgent priority than generalising across backbones, since it would isolate which specific representational choice the two systems make is actually responsible for any performance difference between them.

Finally, and most importantly for how this preprint should be read: this is a first public report on a new architecture, not a final or fully audited result. We are releasing it at this stage specifically to put the central idea, that agent memory can be modelled as a predictive system updated by surprise rather than as a fact store, on the record, and we intend to follow it with ablations, a second benchmark, broader backbone coverage, and a controlled head-to-head comparison against BeliefMem.

8 Conclusion↩︎

We have presented Nous, an agent memory system that treats knowledge as a predictive world model rather than a database of facts. By representing each entity-attribute pair as a categorical probability distribution and updating it via information-theoretic surprise and closed-form Bayesian updates, Nous handles contradiction, forgetting, and multi-value aggregation in a unified framework with roots in cognitive science and information theory.

On the LoCoMo benchmark under GPT-4o-mini, Nous reaches F1 of 63.50, 55.32, 58.57, and 62.50 on single-hop, multi-hop, temporal, and open-domain questions respectively, and outperforms A-MEM’s self-reported numbers on this same backbone, with the caveats on category-mapping reproducibility discussed in Section 5. On a direct comparison against BeliefMem [3], a concurrently developed system built on a closely related belief-based premise, Nous’s self-reported numbers exceed BeliefMem’s self-reported numbers on all four categories, though we treat this as suggestive rather than conclusive pending a controlled, self-run comparison under matched pipelines. The 86 percent single-hop context recall confirms that the Bayesian retrieval pipeline reliably surfaces relevant information; the primary remaining bottleneck for that category is LLM answer extraction rather than retrieval. We present these results as an early but architecturally novel result, with ablations, a second benchmark, broader backbone coverage, and a controlled BeliefMem comparison planned as immediate next steps. All code and the evaluation harness are released publicly.

References↩︎