Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rashomon Memory: Towards Argumentation-Driven Retrieval for Multi-Perspective Agent Memory

Albert Sadowski, Jarosław A. Chudziak

Abstract

AI agents operating over extended time horizons accumulate experiences that serve multiple concurrent goals, and must often maintain conflicting interpretations of the same events. A concession during a client negotiation encodes as a ``trust-building investment'' for one strategic goal and a ``contractual liability'' for another. Current memory architectures assume a single correct encoding, or at best support multiple views over unified storage. We propose Rashomon Memory: an architecture where parallel goal-conditioned agents encode experiences according to their priorities and negotiate at query time through argumentation. Each perspective maintains its own ontology and knowledge graph. At retrieval, perspectives propose interpretations, critique each other's proposals using asymmetric domain knowledge, and Dung's argumentation semantics determines which proposals survive. The resulting attack graph is itself an explanation: it records which interpretation was selected, which alternatives were considered, and on what grounds they were rejected. We present a proof-of-concept showing that retrieval modes (selection, composition, conflict surfacing) emerge from attack graph topology, and that the conflict surfacing mode, where the system reports genuine disagreement rather than forcing resolution, lets decision-makers see the underlying interpretive conflict directly.

Rashomon Memory: Towards Argumentation-Driven Retrieval for Multi-Perspective Agent Memory

Abstract

AI agents operating over extended time horizons accumulate experiences that serve multiple concurrent goals, and must often maintain conflicting interpretations of the same events. A concession during a client negotiation encodes as a ``trust-building investment'' for one strategic goal and a ``contractual liability'' for another. Current memory architectures assume a single correct encoding, or at best support multiple views over unified storage. We propose Rashomon Memory: an architecture where parallel goal-conditioned agents encode experiences according to their priorities and negotiate at query time through argumentation. Each perspective maintains its own ontology and knowledge graph. At retrieval, perspectives propose interpretations, critique each other's proposals using asymmetric domain knowledge, and Dung's argumentation semantics determines which proposals survive. The resulting attack graph is itself an explanation: it records which interpretation was selected, which alternatives were considered, and on what grounds they were rejected. We present a proof-of-concept showing that retrieval modes (selection, composition, conflict surfacing) emerge from attack graph topology, and that the conflict surfacing mode, where the system reports genuine disagreement rather than forcing resolution, lets decision-makers see the underlying interpretive conflict directly.

Paper Structure

This paper contains 20 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Rashomon Memory: Architecture and Components
  3. Observation Buffer
  4. Goal-Perspectives
  5. Ontology Construction.
  6. Knowledge Graph Curation.
  7. Perspective Boundaries.
  8. Retrieval Arbiter
  9. Example: Query Walkthrough.
  10. Retrieval Modes
  11. Case Study: Implementation and Demonstration
  12. Goal-Conditioned Encoding.
  13. Retrieval Through Argumentation.
  14. User-Facing Explanation.
  15. Observations and Limitations.
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Parallel encodings of the concession event. Each perspective preserves what matters for its goals and discards what would interfere.
  • Figure 2: The Rashomon Memory architecture. Raw experiences enter a shared Observation Buffer, where parallel Goal-Perspectives encode them through distinct ontologies into separate knowledge graphs. The Retrieval Arbiter negotiates among perspectives at query time through argumentation, surfacing interpretations that fit the query context, along with an account of what was rejected and why.
  • Figure 3: Attack graphs for four queries. Arrows denote attacks. Blue nodes belong to the grounded extension (accepted); gray nodes are defeated. (a) Risk and Financial attack Relationship; survivors compose. (b) No attacks; all compose. (c) Risk Management attacks both and is sole survivor (selection). (d) Complete mutual attacks; empty grounded extension (surfacing).