Rapid memory
for AI agents.

FishMem is the hosted memory platform built on the open-source rapmem engine. A drop-in mem0-compatible REST API with hybrid graph + vector recall, bi-temporal facts, and ~6-9× faster search — running on the edge, with API keys in minutes.

Get Your API Key Read the Docs

mem0-compatible API
~420ms p50 search
Open-source engine

Quickstart

Set up in minutes

Two REST calls give your agent long-term memory: add what the user says, then search it back whenever you need it.

Create an API key

Open the dashboard → API Keys, create a key, and copy the secret — it starts with rm_ and is shown once.

Add and search memories

POST /v1/memories stores what the user said. POST /v1/memories/search recalls the relevant facts before your next model call. No SDK required — it is just HTTP.

Get your API key

const headers = {
  Authorization: "Bearer rm_your_api_key",
  "Content-Type": "application/json"
};

// 1. Add what the user tells you
await fetch("https://fishmem.com/v1/memories", {
  method: "POST",
  headers,
  body: JSON.stringify({
    messages: [
      { role: "user", content: "I'm vegetarian and allergic to peanuts." }
    ],
    user_id: "alex"
  })
});

// 2. Recall it before the next model call
const response = await fetch("https://fishmem.com/v1/memories/search", {
  method: "POST",
  headers,
  body: JSON.stringify({ query: "What can Alex eat?", user_id: "alex" })
});

const { results } = await response.json();

Memory your agents can actually rely on in production.

The goal is not to hoard transcripts. The goal is to extract the facts that matter, keep them current, and return the right ones in well under a second — every turn of the loop.

Hybrid retrieval

Four signals, one ranked recall

~420ms p50

Vector + keywordMatched

Graph diffusionRanked

Temporal validityResolved

Recall That Understands Time and Relationships

Facts are bi-temporal — they know when they were true and what superseded them — and graph recall connects related facts across conversations, so the right memory surfaces even when the words don't match.

62.2% vs 54.5% LLM-judge accuracy vs mem0 OSS on the held-out set
69.8% vs 28.3% on temporal questions — facts know when they were true
Superseded facts are versioned and stay queryable, never silently deleted

See how retrieval works

Benchmarked on a held-out set, methodology open-sourced

62.2%

POST /v1/memories

Endpoint

/v1/memories

Auth

API key

Mode

Inferred

Status

201

Drop-In mem0-Compatible API

POST /v1/memories to add, POST /v1/memories/search to recall. The REST API matches the shapes your mem0 client already sends, so migrating is a base-URL change.

API keys in minutes, no card required
Raw or inferred adds — the pipeline extracts facts for you
Keep your existing mem0 client and SDKs

Read the API docs

Compatible with your mem0 client

LiveHosted memory platform

Scales automatically and stays fast from every region

Isolated by workspace, exportable via the API — your data stays yours.

Nothing to Operate — Sign Up, Create a Key, Ship

FishMem is memory out of the box. No servers to size, no indexes to babysit, no pipelines to deploy. It scales with you automatically and stays fast from every region.

Live in minutes — sign up, create a key, make your first call
Scales with your traffic automatically, fast from every region
Your data stays isolated, exportable, and deletable at any time

Get your API key Open-source engine

Built on the open-source rapmem engine

OSS

Benchmarks

Measured head-to-head, in the open

rapmem is compared against mem0 OSS on the same harness, with the methodology, prompts, and judge open-sourced so you can rerun it.

LLM-judge accuracy, 0–100%

rapmemmem0 OSS

54.5%

62.2%

Overall

28.3%

69.8%

Temporal

+41.5pt

57.1%

54.8%

Multi-hop

mem0 ahead — published anyway

Use cases

Memory that adapts to your domain

The same two endpoints power very different agents. Scope memories per user, agent, or session, and rapmem keeps each domain's facts current as the world changes.

Customer Support

Agents that remember every prior ticket, the customer's preferences, and the promises your team made — so nobody has to repeat themselves.

AI Companions

Relationships that deepen over time. Companions recall months of conversation and know what changed and when — not just what was said.

Coding Agents

Project conventions, past decisions, and hard-won gotchas persist across sessions, so your agent stops re-learning the codebase every morning.

Sales & CRM

Every conversation, objection, and follow-up stays in context. The next call starts where the last one ended, not from a blank page.

Healthcare Assistants

History, allergies, and what helped before — with full change history per fact, so the compliant workflows you build on top stay grounded in accurate, current records.

Education

Tracks what each learner has mastered and where they struggle, so tutoring agents pick up exactly where the student left off.

Built for teams. Designed for control.

No black boxes around your users' memories. Isolation is enforced in the data layer, every change is logged, and your data is exportable on demand.

Get your API key

Isolated

Every workspace's memories are hard-isolated in the data layer — no cross-tenant reads, no shared indexes. API keys are hashed at rest, shown once at creation, and revocable instantly from the dashboard.

Auditable

Every add, update, invalidate, and delete is logged per memory, and the full change history is one API call away. Usage is metered per operation, so you always know what ran and what it cost.

Yours

Export or delete everything via the API at any time. The engine is open source, so there is no lock-in — and self-hosting is always an option.

Start free, upgrade when traffic grows

Prototype on Hobby, then move to monthly Stripe plans with higher included memory operations. Reads are always free.

Get your API key

Hobby

Freeno card required

Get your API key

Credits: 10,000 add requests and 1,000 retrieval requests per month
Best for: Prototyping memory for a single agent or side project
Metering: Full API access from day one: adds, search, and free reads
Support: Community support on GitHub issues and discussions
Platform access: mem0-compatible REST API, dashboard, and one workspace
Data ownership: Memories stay workspace-scoped; export or delete them at any time.

Starter

Growth

$79per month

Start Growth

Credits: 200,000 add requests and 20,000 retrieval requests per month
Best for: Teams scaling memory across production users
Metering: Usage analytics, API keys, webhooks, and billing controls
Support: Email support from the engineers who built the engine
Platform access: Multiple projects, hybrid retrieval, graph and temporal recall
Data ownership: Memories stay workspace-scoped; export or delete them at any time.

Pro

$249per month

Start Pro

Credits: 500,000 add requests and 50,000 retrieval requests per month
Best for: Higher-volume apps that need advanced usage visibility
Metering: Advanced analytics and higher included monthly capacity
Support: Private Slack support
Platform access: Unlimited projects with production-ready controls
Data ownership: Memories stay workspace-scoped; export or delete them at any time.

Enterprise

Customannual commitment

Talk to us

Credits: Committed-use credit pricing with volume discounts
Best for: Teams with compliance, scale, or data-residency requirements
Metering: Dedicated capacity, SLAs, and priority on the roadmap
Support: Private channel, onboarding help, and migration assistance
Platform access: Everything in Pro, plus SSO and audit controls
Data ownership: Memories stay workspace-scoped; export or delete them at any time.

Plan	Hobby Freeno card required Get your API key	StarterMost popular $19per month Start Starter	Growth $79per month Start Growth	Pro $249per month Start Pro	Enterprise Customannual commitment Talk to us
Credits	10,000 add requests and 1,000 retrieval requests per month	50,000 add requests and 5,000 retrieval requests per month	200,000 add requests and 20,000 retrieval requests per month	500,000 add requests and 50,000 retrieval requests per month	Committed-use credit pricing with volume discounts
Best for	Prototyping memory for a single agent or side project	Small apps moving from prototype to live traffic	Teams scaling memory across production users	Higher-volume apps that need advanced usage visibility	Teams with compliance, scale, or data-residency requirements
Metering	Full API access from day one: adds, search, and free reads	Operation-level usage and credit balance in the dashboard	Usage analytics, API keys, webhooks, and billing controls	Advanced analytics and higher included monthly capacity	Dedicated capacity, SLAs, and priority on the roadmap
Support	Community support on GitHub issues and discussions	Community support	Email support from the engineers who built the engine	Private Slack support	Private channel, onboarding help, and migration assistance
Platform access	mem0-compatible REST API, dashboard, and one workspace	Hosted FishMem API, dashboard, and one workspace	Multiple projects, hybrid retrieval, graph and temporal recall	Unlimited projects with production-ready controls	Everything in Pro, plus SSO and audit controls
Data ownership	Memories stay workspace-scoped; export or delete them at any time.	Memories stay workspace-scoped; export or delete them at any time.	Memories stay workspace-scoped; export or delete them at any time.	Memories stay workspace-scoped; export or delete them at any time.	Memories stay workspace-scoped; export or delete them at any time.

Credits are consumed per API operation: 2 for an inferred add, 1 for a raw add or search, and reads are free. Self-hosting the open-source engine costs nothing.

Honest numbers. No benchmark theater.

“We publish the harness, the prompts, and the judge. 62.2% vs 54.5% on LOCOMO means nothing unless you can rerun it yourself.”

Honest benchmarks

Evaluation principle

“A memory store that forgets when facts were true will confidently answer with stale ones. Bi-temporal tracking is not optional.”

Time-aware truth

Data model principle

“Memory sits inside the agent loop, so latency compounds. A ~420ms p50 search is a feature, not an implementation detail.”

Speed as a feature

Performance principle

Frequently asked questions

Quick answers about FishMem, the rapmem engine, and migrating from mem0.

What is FishMem?

FishMem is a hosted memory layer for AI agents. It runs on the open-source rapmem engine, extracts durable facts from conversations, stores them in a bi-temporal knowledge graph, and recalls them with hybrid vector, keyword, temporal, and graph-diffusion retrieval.

How is FishMem different from mem0?

The API is mem0-compatible, but the recall is different: hybrid retrieval with Personalized-PageRank graph diffusion, bi-temporal facts that know when they were true, and profile blocks. On held-out LOCOMO conversations rapmem scores 62.2% vs 54.5% LLM-judge accuracy against mem0 OSS (+41.5pt on temporal questions), and search is ~6-9× faster (~419ms p50 vs 2.4-3.8s). The comparison ran on the same harness, and the methodology is open-sourced.

Can I migrate from mem0 without rewriting my code?

Yes. POST /v1/memories and POST /v1/memories/search accept the request shapes your existing mem0 client sends. In most cases migration is changing the base URL and the API key — your add and search calls keep working.

How is my data isolated and secured?

Every memory is scoped to a workspace, and isolation is enforced in the data layer — no cross-tenant reads or shared indexes. API keys are shown once at creation and stored as salted hashes, and you can rotate or revoke them from the dashboard. You can export or delete a workspace's memories at any time.

What is the difference between the open source engine and the cloud?

The rapmem engine is open source (github.com/rapmem/rapmem) under Apache-2.0, and you can self-host it for free. FishMem runs the same engine and adds the dashboard, API key management, metering, and billing — you get keys in minutes instead of operating memory infrastructure yourself.

How does pricing work?

Usage is metered in transparent credits: an inferred add is 2 credits, a raw add is 1, a search is 1, and reads and lists are free. Hobby is free, and Starter, Growth, and Pro are monthly Stripe subscriptions with larger included credit buckets.

Rapid memory for AI agents.

Set up in minutes

Create an API key

Add and search memories

Memory your agents can actually rely on in production.

Recall That Understands Time and Relationships

Drop-In mem0-Compatible API

Nothing to Operate — Sign Up, Create a Key, Ship

Measured head-to-head, in the open

Memory that adapts to your domain

Customer Support

AI Companions

Coding Agents

Sales & CRM

Healthcare Assistants

Education

Built for teams. Designed for control.

Isolated

Auditable

Yours

Start free, upgrade when traffic grows

Hobby

Starter

Growth

Pro

Enterprise

Honest numbers. No benchmark theater.

Frequently asked questions

Rapid memory
for AI agents.