How to choose an AI memory layer that forgets correctly

Most AI agents just re-read a transcript. A technical analysis of MemPalace, Zep, Mem0, and others reveals the critical, unsolved problem of how agents should forget outdated information.

For developers building stateful AI agents, the default memory solution of stuffing the conversation transcript into the context window is a liability. It's expensive, noisy, and dangerously naive about outdated facts. If your agent needs to remember a user's peanut allergy but forget their old shipping address, you need a real memory layer. For simple, stateless bots or prototypes where conversations are short, transcript stuffing is adequate. But for any serious application, the core challenge is not remembering more, it is forgetting correctly and safely. The best architecture separates hot, warm, and cold storage and treats fact invalidation as a first-class operation.

Methodology

This is a v0 review analyzing the concept of AI agent memory architecture. It draws exclusively on the technical analysis published by Diran Adeola on June 15, 2026. The source signal is a blog post titled "Your AI agent doesn't have a memory. It has a transcript," available at https://dev.to/diran_adeola/your-ai-agent-doesnt-have-a-memory-it-has-a-transcript-16lp.

This review covers the author's critique of transcript-stuffing and his comparative analysis of five existing memory tools: MemPalace, Zep, Obsidian, Mem0, and Letta/MemGPT. The focus is on the conceptual models these tools represent, particularly regarding data decay and retrieval. What is not covered are independent performance benchmarks, hands-on implementation details of these tools, or a full review of the author's proposed alternative architecture. All characterizations of the five tools are based on the author's public claims; independent benchmarks are pending. We will re-evaluate when new performance data becomes available.

What it does

The source post dissects the failure modes of naive AI memory and surveys the current landscape of dedicated solutions.

The problem with transcript stuffing

According to Adeola, relying on the conversation transcript as a memory system fails in three distinct ways. First, it's expensive and noisy. Every turn requires re-sending the entire history, burying critical facts like "I'm allergic to peanuts" in conversational filler, with costs scaling linearly. Second, it cannot handle stale data. If a user states "I'm vegetarian" in March and "I eat fish now" in May, both facts are presented to the model with equal weight, forcing it to guess. Third, when summarization is added to control costs, it can silently drop critical information. The author frames this not as a cost bug, but as a safety bug.

A survey of existing memory layers

The author analyzes five tools, concluding each solves one part of the problem well but is weak elsewhere, especially in managing how information is forgotten or invalidated.

• MemPalace: Never discards data, keeping everything in the prompt. This provides a safety net but fails to address cost or context window limits.
• Zep: Correctly models fact validity with a valid_until field, making superseding facts a core concept. Its main drawback is the claimed operational weight of running a graph database.
• Obsidian: Excels at creating a human-readable, version-controlled knowledge layer in Markdown. It lacks semantic recall and automated data decay.
• Mem0: Focuses on compact, typed data extraction. The author claims this process occurs on the hot path, making writes potentially expensive.
• Letta / MemGPT: Treats the context window like OS RAM, paging memory in and out. The author finds its placement of paging logic within the agent's reasoning to be a strange architectural choice.

What's interesting / What's not

The most interesting contribution here is the reframing of the agent memory problem. The shift in focus from "how to remember more" to "how to forget correctly" is the critical insight for anyone building robust agents. Adeola’s argument that dropping an allergy to save tokens is a safety failure, not a cost optimization, is a powerful and necessary perspective for the field.

The taxonomy of existing tools is also highly valuable. It provides a clear mental model for developers to map their own needs against the strengths of available solutions. Zep's explicit fact validity is a strong pattern, while Obsidian's human-readable layer is compelling for different reasons. The author's stated principle, "be deeply unoriginal in your parts and opinionated in your assembly," is a pragmatic engineering philosophy.

What's missing is concrete data. The analysis is conceptual, and the author's proposed three-tier architecture remains a mental model without a public implementation or benchmarks. The critiques of the five tools, while architecturally sound, are not backed by performance numbers. For instance, the claim that Mem0's writes are expensive or that Zep's graph database is too heavy is presented without a specific test case or workload.

Pricing

Pricing is not discussed in the source analysis and varies significantly across the tools mentioned. Zep offers both an open-source, self-hosted version and a managed cloud service with usage-based pricing. Obsidian is a personal knowledge management tool with a free personal license and a commercial license for business use. Mem0 and MemPalace appear to be open-source projects. Readers should consult each project's individual pricing page for current information. (Snapshot: June 15, 2026).

Verdict

The immediate takeaway for builders is that stateful AI agents require a memory system that explicitly handles fact invalidation and tiered storage. Simple transcript stuffing is a liability for any production system. For teams evaluating options, the choice depends on specific constraints. If you can manage the operational overhead, Zep's approach to fact validity is architecturally sound. If human-readability and versioning are paramount, an Obsidian-like knowledge base with a semantic search layer is a viable path. The author's proposed hybrid model is compelling, but remains conceptual. The most important action is to stop treating memory as an ever-growing transcript and start designing for data decay.

What we'd test next

A v2 review would require hands-on implementation. We would construct a reference agent and a long-running conversation that includes fact updates, retractions, and critical information. We would then implement the memory layer using at least two of the discussed tools, likely Zep and Mem0, alongside a simplified version of the author's proposed tiered architecture. Key metrics for comparison would include cost per API call as history grows, latency for memory retrieval, accuracy in retrieving the most current fact when conflicting information exists, and a qualitative assessment of how each system prevents the loss of critical facts during summarization or compaction.

The investor read

This analysis signals a maturation of the AI agent market. First-generation 'transcript stuffing' is giving way to specialized, robust memory layers, creating a new infrastructure category. The critical insight is that agent memory is a state management problem where forgetting is a feature, not a bug. Zep is a notable commercial player here. Investable companies will likely offer a managed service that abstracts away the operational complexity of these systems (as Zep's cloud offering does) while providing strong guarantees around data consistency and safety. The market is shifting from generic vector databases to purpose-built memory systems. A company that can productize a clean, tiered architecture for decaying memory could capture significant developer spend.

Pull quote: “The core challenge is not remembering more, it is forgetting correctly and safely.”