There is no evidence-backed universal winner among persistent memory APIs for AI agents. The right choice depends on what the system remembers, how those memories are updated and shared, who controls them, and where the service runs. Start with a shortlist based on those differences, then compare candidates on the same representative tasks.
What a persistent memory API actually decides
Persistent memory is more than a storage endpoint. A system has to decide what to extract from conversations or other inputs, how to represent it, when to retrieve it, and how to revise or remove it. Those choices affect the behavior of the agent as much as the underlying storage does.
As an Amazon Associate I earn from qualifying purchases.
For example, an API that turns messages into extracted facts gives the service an active role in deciding what to retain. An agent-managed approach instead gives the agent tools to rewrite its own memory. A temporal graph or profile-oriented system makes different representation choices. These are distinct operating models, not interchangeable database interfaces.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Which approaches belong on a shortlist?
The available comparison is useful as a taxonomy, not as an independently verified ranking. It was published by Dosu on September 15, 2026, and its provider-authored descriptions should be checked against each product’s current documentation before a purchase or implementation decision.
| Candidate or approach | What the available material describes | What to validate |
|---|---|---|
| Mem0 | Its public quick-start documents adding messages with a user ID and searching with a filter on that same ID. The product describes itself as a persistent memory layer for agents. | Whether user-scoped add-and-search fits your memory policy; how extraction, updates, corrections, deletion, sharing, and deployment work for your use case. |
| Zep | Zep positions itself as an enterprise context layer with agent-memory capabilities. Its current product page describes a Memory MCP Server intended to provide shared memory for each user across agents, governed by policy. | How the documented policy controls behave, which agents and clients can access shared memory, and the exact API and deployment behavior you need. |
| Supermemory | The Dosu comparison describes a combination of extraction, profiles, and document retrieval. | How those components interact, what is retained or retrieved, and whether the current product documentation supports your required controls. |
| Letta | The comparison describes an agent-managed approach in which the agent has tools to rewrite its memory. | What guardrails govern those writes, how changes are audited or reversed, and how memory is shared across agents. |
| LangMem | The comparison describes a library that packages memory tools for agent frameworks. | How much of the memory policy and infrastructure your team must implement and operate. |
| Redis Agent Memory or Postgres with pgvector | The comparison lists these as alternatives that allocate more implementation responsibility to the builder. | Which extraction, update, retrieval, access-control, and deletion behaviors you must build yourself, and whether that control is worth the operational work. |
These descriptions reflect the cited comparison’s characterization; they do not establish feature parity, current availability, or comparative performance. Treat each row as a reason to investigate, not as a substitute for checking current provider documentation.
How to compare memory models
Extraction and representation
Ask what becomes a memory: explicit user preferences, inferred facts, conversation episodes, documents, or some combination. Find out whether the system stores original inputs, derived records, or both, and how it handles contradictory or outdated information. A system that extracts facts may reduce the work of building a memory pipeline, but its extraction decisions also need to match your application’s standards.
Updates, correction, and forgetting
Test what happens when a user corrects a fact, changes a preference, asks for something to be forgotten, or provides conflicting information. Establish whether updates replace, merge, or coexist with older records; whether deletion reaches derived memories as well as source records; and how you can confirm that an item is no longer retrievable. Do not assume that a general-purpose delete operation covers every representation or backup.
Sharing across agents and clients
Determine whether memory belongs to a user, an agent, a session, or another scope, and how access is authorized. Mem0’s documented quick-start illustrates user-ID filtering for its basic add-and-search workflow. Zep’s product description presents its Memory MCP Server as a way to share each user’s memory across agents under policy. Those descriptions point to different questions about scope and governance; they are not proof that either approach meets a particular access-control requirement.
Deployment and governance
Compare hosted and self-managed options using your own requirements for data handling, access, retention, auditability, and operational ownership. The evidence available here does not establish a current cross-provider security or compliance matrix, regional availability, or relative pricing. Verify those details directly in current vendor materials before procurement rather than inferring them from product positioning.
Integration effort
A managed memory layer can take on extraction and retrieval work, while a framework library or database-oriented approach may leave more of that design and operations to your team. Include the full integration burden in the comparison: data preparation, policies, monitoring, failure handling, and maintaining behavior as models and prompts change—not just the number of API calls required to get a prototype running.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to evaluate candidates on your own tasks
Run a controlled comparison using a representative, privacy-safe sample of conversations and tasks. Decide in advance what counts as a useful memory, what must never be stored, and how you will score results. Keep the generation model, embedding model, extraction prompts, dataset, and retrieval settings consistent across candidates whenever possible.
Recommended Free Tools
- Define memory rules. Specify what the system may remember, what is prohibited, how long information should persist, and what should happen when facts become stale or conflict.
- Build test scenarios. Include useful recall, irrelevant recall, corrections, changed preferences, explicit forgetting requests, and access from the agents or clients that should—and should not—share a user’s information.
- Test writes and reads. Record whether each candidate stores the intended information and returns relevant memories for the same questions. Check for false, stale, or sensitive recall as well as missed recall.
- Verify lifecycle controls. After changing or deleting information, query again to check whether the old memory is still returned. Inspect the provider’s documented behavior for derived records and other retention paths relevant to your requirements.
- Measure operational trade-offs. Track read and write latency, operational cost, integration work, and the reliability of correction and deletion alongside retrieval usefulness.
- Repeat under fixed conditions. Use the same models, prompts, data, and retrieval configuration for each candidate; record any unavoidable differences so they do not get mistaken for product effects.
This evaluation plan is a practical way to apply the documented differences; it is not a published benchmark or a claim that these providers have been tested under identical conditions.
Why benchmark scores do not settle the choice
The September 15, 2026 Dosu comparison cautions that results on LongMemEval and LoCoMo can depend on the generation model, embedding model, extraction prompts, and retrieval settings. It also reports that Mem0 and Zep have publicly disputed one another’s reported results. A score produced with one configuration therefore does not establish a universal best API. Compare methodology as well as the headline number, and give greater weight to results that reproduce your application’s workload.
Is there a shared standard for memory APIs?
A June 2026 memorywire preprint proposes a vendor-neutral JSON Schema for five operations—remember, recall, forget, merge, and expire—and four memory types—semantic, episodic, procedural, and emotional. That is evidence of interoperability work, not an adopted industry standard. Do not assume that similarly named operations behave consistently across providers or that a memory created for one system can be moved to another without translation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
