Skip to content

From Agent Memory to Enterprise Memory

Executive summary

  • Memory can improve agent performance. Enterprise value depends on retaining knowledge that is correct, current and permitted for reuse.
  • Obsidian can provide a readable knowledge workspace; a temporal graph can connect facts, sources and changes over time.
  • Zep is the first candidate to evaluate for the architecture below. Measure complete operating cost and accepted outcomes before scaling.
5 minute readExplore the memory lifecycle
From Agent Memory to Enterprise Memory

An AI agent can finish a task and still leave the company no wiser.

Enterprise memory is how useful experience becomes reusable knowledge.

What Anthropic's research establishes

In its 6 November research, Anthropic describes agents retaining experience between sessions and reorganising it into more useful memory. This is external context, rather than retraining the model.

On the CL-Bench database task, it reports mean reward rising from 0.655 to 0.838 for Opus 5.5 when a persistent memory store was available. Each question started in a fresh session, with feedback available to inform later answers.

That is a 0.183 increase on the benchmark's reward scale. It is evidence of improvement on that task, not a measured increase in enterprise productivity.

Another finding matters just as much: some agents accumulated untested explanations, while stronger approaches recorded checked rules and corrections. Anthropic research

The enterprise implication is clear: writing something to memory must not automatically make it company knowledge.

Observed vendor benchmark: Opus 5.5 on the CL-Bench database task. Mean reward uses a zero-to-one scale. This is not a productivity or ROI measurement.

Anthropic research, 6 November 2026
What the research measured
Accessible chart data / same model, same task
ConditionMean reward (0 to 1)
Stateless0.655
Memory store0.838
Difference in reward points+0.183

No uncertainty intervals or run costs were available in the accessed figure description.

Where Obsidian fits

Obsidian stores notes as local Markdown files. Its graph displays links between those notes. This makes it a useful workspace for inspecting decisions, procedures and their supporting evidence. Storage documentation · Graph documentation

The proposed design separates three responsibilities:

Three responsibilities in the proposed design
LayerPurpose
Source systems and data lakeRetain authoritative records and their history.
Obsidian or another knowledge workspaceLet authorised people review and maintain approved knowledge.
Governed memory serviceRetrieve relevant facts, relationships and evidence for each agent task.

A note-link graph is different from a temporal knowledge graph that represents customers, contracts, obligations and when each relationship applies.

An Obsidian integration would require an ingestion and approval workflow. A shared vault is not the access-control boundary for every employee and agent.

The candidate worth evaluating first

For this enterprise design, Zep is the strongest first candidate among the options reviewed. The reason is architectural fit: changing relationships, source provenance and managed context delivery.

Graphiti is its open-source temporal graph framework. It tracks changing facts and their source episodes. Self-hosting Graphiti means taking responsibility for the database, ingestion, access control and operations. Zep adds a managed enterprise context service; the two should not be treated as equivalent deployments. Graphiti documentation

Zep reports 451 correct answers out of 500 on LongMemEval, with 4,408 median context tokens and 162 ms p95 retrieval latency. Its setup uses GPT-5.4 for answering and judging. These are vendor results for that configuration, not independent enterprise outcomes or measurements of a standalone Graphiti installation. Retrieval latency also excludes the time needed to generate the final answer. Benchmark methodology

Mem0 and Hindsight remain credible challengers. Mem0 documents fact extraction and retrieval; Hindsight explicitly separates retaining, recalling and reflecting on experience. Their published scores are not a controlled comparison with Zep. Mem0 research · Hindsight implementation

The design decision that creates value

Consider a delivery agent preparing a customer handover. A meeting suggests a configuration change; an approved change record later confirms it.

The memory system should preserve the distinction between suggested, approved and superseded. It should retrieve the applicable decision and its evidence, rather than whichever note is most recent or most similar to the question.

The proposed loop is:

Capture → validate → approve for reuse → retrieve → observe the outcome → revise.

Each reusable item needs an owner, source, scope, validity period and permitted audience. New evidence can trigger revalidation. Current permissions must be checked before retrieval, and source deletion or revocation must propagate to derived memory.

This makes IP-In/IP-Out operational: authorised knowledge enters a task; validated learning can leave it as reusable IP when ownership and permissions allow.

Proposed architecture. Source systems remain authoritative. A trusted approval and ingestion service publishes governed memory; agents cannot approve their own memories.

Original analysis by João Frederico Neto. This diagram is a proposal, not a tested integration.

How enterprise memory earns reuse

Prove the benefit on your own work

Compare the same model on the same held-out tasks using existing document retrieval, curated Markdown and temporal memory. Include changed decisions, conflicting sources and revoked access.

Measure answer correctness, stale-fact errors, human correction time, end-to-end latency and total cost per accepted task. Include extraction, storage, retrieval, model usage and knowledge maintenance.

The potential benefits are less repeated discovery, better handovers and fewer decisions based on obsolete information. Public benchmarks justify testing those benefits; they do not establish your savings.

Build memory that earns reuse, then measure whether the business benefits from remembering.

THE INTERACTIVE COMPANION

What should the agent
be allowed to remember?

Explore six evidence states and see when a memory can support an answer.

Explore the memory lifecycle
More from Our thinking