Distilling Agent Episodes into Provenance-Bearing Graphs

#Abstract

An agent that completes a task holds a transcript, and a transcript has no operation that returns the fact that two of them disagree. This paper describes a construction in which an episode is compressed into typed claims that retain the observations supporting them, written as graph edges over content-addressed nodes, so that later contradiction becomes a traversal. Revision cost is bounded by the transitive support closure. The confidence composition rule assumes independence, and the paper is explicit that this assumption fails in a common case.

#Why transcripts are the wrong long-term store

Three properties, all structural rather than incidental.

They do not compose. Two episodes concerning one entity produce two internally coherent transcripts with nothing linking the entity across them.
They have no contradiction surface. Disagreement is visible only once both are read into a context window by a model that may or may not notice, and attention to material in long inputs is known to be position dependent .
They price recall by length. Retrieval cost grows with history rather than relevance.

Retrieval augmentation addresses the third by fetching selectively and does not address the first two, because the retrieved unit is still text.

#The distillation step

At episode close, each conclusion is emitted as a claim carrying the observations it rests on and a confidence.

c \;=\; \langle s,\; p,\; o,\; \Pi(c),\; \kappa(c) \rangle, \qquad \Pi(c) \subseteq O_{e}

(1)

A claim, with the support set that makes it revisable

Π(c) is the support set from episode e, and κ(c) the confidence. A claim with an empty support set is malformed rather than weak and is rejected by the distiller. That single rule removes a substantial proportion of low-quality output before it reaches the graph, because a model asked to summarise will readily produce statements it cannot ground .

#Identity

Revision requires that two episodes mentioning one entity produce edges on one node. Node identity is derived as a hash of tenant, kind and normalised external identifier rather than allocated , so convergence is structural rather than achieved by a matching pass. Multi-source identity and its failure modes are treated separately.

#Confidence across episodes

\kappa(c) \;=\; 1 - \prod_{i=1}^{n}\left(1 - \kappa_i(c)\right)

(2)

Composition over support sets assumed independent

Agreement between independent observations should count for more than either alone, which this expresses.

The independence assumption is the weakest step in the construction. Two episodes that both consulted the same upstream document are not independent, and the rule will overstate confidence. Treating source documents as first-class nodes makes shared provenance detectable, and detecting it is not the same as calibrating for it. No calibration is offered.

#Revision as traversal

R(o) \;=\; \bigl|\{\, c \in C \;:\; o \in \Pi^{*}(c) \,\}\bigr| \;\ll\; |C|

(3)

Cost of revising a retracted observation

The transitive support closure Π* bounds the work. Without retained provenance the only sound response to a retraction is full re-derivation, which is the behaviour transcript-based memory falls back to, usually by not responding to retractions at all.

#Relation to existing agent memory

The generative agents architecture maintains a memory stream retrieved by recency, importance and relevance, with periodic reflection producing higher-level statements . The construction here is close in shape, and differs in that reflections are stored as graph edges with retained provenance rather than as text, which is what makes them revisable.

Graph-structured retrieval has separately been shown to support query-focused summarisation over corpora where flat retrieval performs poorly , and the representational arguments for graphs over text apply here for the same reasons .

#Limitations

Independence, as above. This is the defect a reviewer will find first and it is acknowledged rather than resolved.

Selection. Which conclusions become claims is decided by a model. A conclusion never emitted is silently lost with no signal at the time. Emitting low-confidence claims liberally and pruning later trades a silent-loss problem for a graph-pollution problem, and no evidence is offered on which is preferable.

Calibration across models. Confidence values from models of differing verbosity are not on a common scale, and the composition rule assumes they are.

No end-to-end evaluation. No measurement is reported showing that an agent using this construction answers better than one using a transcript with retrieval. That experiment is the one that would establish the practical claim.

References

[1]Joon Sung Park et al., “Generative Agents: Interactive Simulacra of Human Behavior”, ACM Symposium on User Interface Software and Technology (UIST), 2023doi:10.1145/3586183.3606763 ↗
[2]Nelson F. Liu et al., “Lost in the Middle: How Language Models Use Long Contexts”, Transactions of the Association for Computational Linguistics, vol. 12, pp. 157-173, 2024doi:10.1162/tacl_a_00638 ↗
[3]Aidan Hogan et al., “Knowledge Graphs”, ACM Computing Surveys, vol. 54, no. 4, 2021doi:10.1145/3447772 ↗
[4]Ziwei Ji et al., “Survey of Hallucination in Natural Language Generation”, ACM Computing Surveys, vol. 55, no. 12, 2023doi:10.1145/3571730 ↗
[5]Patrick Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, Advances in Neural Information Processing Systems (NeurIPS), 2020
[6]Ralph C. Merkle, “A Digital Signature Based on a Conventional Encryption Function”, Advances in Cryptology (CRYPTO 87), Springer, 1988doi:10.1007/3-540-48184-2_32 ↗
[7]Darren Edge et al., “From Local to Global: A Graph RAG Approach to Query-Focused Summarization”, arXiv:2404.16130, 2024