Skip to content

ResearchPaper

Organizational memory for interchangeable agents

In organizations where AI agents do the work, models and agents are replaced frequently. This paper proposes keeping memory in documents the organization manages rather than inside any model, and turning working records into verified knowledge in stages. It presents a five-stage flow (Trace, Episode, Research, Review, Library) and its invariants, then reports observations from use, a discussion, and limitations.

  • Memory
  • Agents
  • Organizations

Abstract

In organizations that delegate work to AI agents, the entity doing the work changes frequently. Models are updated every few months, and the same task may be handled by models from different providers (for example, Claude, Codex, or Gemini). The organization must nonetheless retain its past decisions and failures and apply them to the next task. This paper proposes placing memory not inside any model but in documents the organization manages. Agents share no internal state; each begins its work by reading documents the organization trusts. As a process for turning working records into verified knowledge in stages, the paper presents five stages: Trace, Episode, Research, Review, and Library.

1. Background

The memory of agents built on large language models is usually confined to a single agent or a single product: a conversation history, summarized memories, or a product’s own storage. This arrangement suits a single agent holding a long conversation with one user. When an organization’s work is carried by several agents and models that replace one another, however, memory is lost along with whoever was doing the work.

For example, an agent that starts fresh in each session will repeat a mistake if it does not know about a correction a person made in the previous session. This has the same structure as the handover problem in human organizations when responsibilities change hands. The difference is that agents change hands far more often than people do.

2. Problem

Keeping memory inside whoever does the work has three weaknesses as a form of organizational memory.

  • Portability. When models or products change, memory is either left behind or has to be converted before it can move.
  • Verification status. Records accumulate, but nothing distinguishes established facts from provisional guesses.
  • Consistency. The same fact is written in several places, and the copies eventually diverge.

The question of this paper is where, and in what form, memory should be kept so that an organization can retain its past decisions and failures, and apply them correctly to the next task, even as those doing the work are replaced.

  • Organizational memory. Walsh and Ungson (1991) defined organizational memory as stored information from an organization’s history that can be brought to bear on present decisions, and showed that it is not retained only in individuals. This paper addresses the case in which the workers are agents rather than people, and proposes documents as the place where that memory is retained.
  • Retrieval-augmented generation. Lewis et al. (2020) showed how a model can retrieve and draw on external documents during generation. The design in this paper likewise assumes that agents read external documents. Its concern, however, is less the retrieval method than the verification status of the documents being retrieved and the process by which those documents come to exist.
  • Memory management within an agent. Packer et al. (2023), drawing on the memory hierarchy of operating systems, showed how memory can be kept outside a limited context and paged in and out. This extends the memory of a single agent. This paper concerns memory held on the organization’s side, independent of any particular agent or model.

4. Design

4.1 Keeping memory in documents

Memory is designed around where something is written, not who remembers it. Every agent reads from the same place when it begins work and writes back to the same place when it finishes. Expertise, too, belongs to places rather than to agents: an agent takes on a role by reading the rules and guides kept in that place.

4.2 A five-stage flow

Records become knowledge through five stages.

  1. Trace. The raw record of what happened: conversations, commands, changes. Stored without modification.
  2. Episode. A single unit of work: what was requested, what was done, and how it ended.
  3. Research. Recurring patterns and hypotheses drawn from multiple episodes.
  4. Review. A person or a separate agent checks the claims made in Research against the original records.
  5. Library. Verified knowledge. This is what an agent reads first on its next task.

4.3 Invariants

The flow is designed to satisfy three conditions.

  • Stage and trust. Each successive stage is smaller in volume and higher in trust.
  • Traceable provenance. Knowledge in the Library can be traced back to the records it was derived from.
  • Single statement. Each fact is written in exactly one place in the Library.

5. Observations from use

Our World is testing this design in Moss, a multi-agent system under research and development, and uses it in the company’s day-to-day work. Two observations have emerged. Both are based on operational experience; no quantitative measurements were made.

First, agents’ errors stem less from how documents are arranged than from inconsistent and missing records. Because agents gather what they need by searching, differences in folder structure were not a major problem. What caused problems were records that contradicted each other and decisions that were never written down. In both cases, the result surfaced as an incorrect answer given to a person.

Second, people need a means of checking what the memory contains. An early prototype kept nearly all records in a database that could be inspected only through a text interface, which made it difficult for people to verify whether the records were correct. The current approach being tested separates material by state: work under consideration in git, decisions in force in a database, and finished material that will not change in long-term storage. A separate interface is provided for people to review it.

6. Discussion

Keeping memory in documents makes it resilient to changes in who does the work. When models or providers change, the documents remain, and any agent can read them. Documents are also readable by people, which allows people to verify what the memory contains, a property that memory inside a model does not have.

This design carries three costs, however. First, if people check everything at the Review stage, people become the bottleneck. Second, it takes time for records to reach the Library, so recent decisions are not reflected immediately. Third, whether an agent finds the right document depends on the quality of its search. On the third point, as the observations from use suggest, preventing inconsistent and missing records may matter more than refining how documents are arranged.

7. Limitations

  • The observations come from small-scale use in a single organization and include no quantitative evaluation.
  • The design has not been compared, under equal conditions, with methods that give models internal memory or with other memory designs.
  • The boundaries between the five stages, particularly the criteria for Research and Review, have not yet been settled in practice.

8. Future work

  • Evaluation. How should one measure whether the memory is working? One candidate is whether an agent, when asked by a person, returns the correct answer together with its source. There is not yet a way to make this a reproducible test.
  • The cost of review. How much of the human checking can be replaced by agents checking one another?
  • Forgetting. When and how should outdated knowledge be retired? Rather than deleting it, the approach under consideration records that it has been superseded.
  • Relation to memory inside models. Does research on giving models internal memory compete with the design in this paper, or complement it?

References

  • Lewis, P., Perez, E., Piktus, A., et al. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33 (NeurIPS 2020).
  • Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., & Gonzalez, J. E. (2023). MemGPT: Towards LLMs as operating systems. arXiv:2310.08560.
  • Walsh, J. P., & Ungson, G. R. (1991). Organizational memory. Academy of Management Review, 16(1), 57–91.

Related

Version history

  1. v0.4Oct 7, 2026Rewritten as a paper. Added related work, the design's invariants, a discussion, and limitations.
  2. v0.3Oct 4, 2026Added what we have learned from internal use and rewrote the open questions.
  3. v0.2Aug 27, 2026Organized the path from records to knowledge into five stages, from Trace to Library.
  4. v0.1Jul 24, 2026First version: the problem, and the choice to keep memory in documents.

This document’s URL will not change. When citing it, please include the version number.

All research