Reflexio Review (2026): Pricing, Features & Honest Verdict

Reflexio is an open-source learning platform that turns user corrections and failed agent runs into scoped, reversible behaviour rules your AI agent reads on the next call. Verdict: the most honest self-improvement tooling we have tested, with a brutal price cliff. Best for: teams running production agents in Python. Price: From $299/mo (free plan: yes). Rating: 7.5/10.

What is Reflexio?

We found Reflexio at number two on Product Hunt on 5 September 2026 with 278 upvotes. It sits in the slot most people call “agent memory”, and the vendor spends a good part of its site explaining why that is the wrong word. Memory writes down what was said. Reflexio writes down what should change. When a user corrects an agent, when a path fails, or when something works well, it extracts a learning, scopes it to a user profile or an agent playbook, and injects only the relevant learnings at the moment of inference. Nothing is retrained. Every learning is readable, editable, approvable and deletable, and a rejected one drops out of retrieval straight away.

The company is Reflexio, Inc. Its terms were last updated on 22 July 2026 and state that customer content is not used to train Reflexio’s own models or models shared between customers. The core is open source under an unmodified Apache 2.0 licence: the main repository was created in April 2026, has 360 stars and 48 forks, and was pushed to on the morning we checked. A companion plugin called claude-smart turns Claude Code, Codex and OpenCode corrections into rules and has 773 stars on its own. This is a real project with active development, not a landing page with a waitlist.

What Are Reflexio’s Key Features?

The publish-and-retrieve loop

Your agent publishes each interaction, meaning the turn, any feedback, actions taken, images and expert examples, then retrieves learned context before the next one. You wire that through a Python SDK, a REST API or the CLI. There is also a portable skill you hand to Codex, Claude Code or Cursor that inspects your agent’s lifecycle, implements the loop and verifies the changed path. If you already run mem0, a drop-in wrapper mirrors hosted mem0 traffic to Reflexio with one import change, so you can compare the two without a rewrite.

Profiles and playbooks

Two scopes carry the learning. User profiles hold durable context about a person and their environment, such as output conventions or production constraints. Playbooks hold procedures: the ordered actions, checks and fallbacks that make a class of task repeatable, at either user or agent scope. Playbooks can be aggregated across users and put through an approval step, and there is an optional GEPA-backed pass that refines the wording of pending rules. Learnings retire when newer conversations contradict them. The site’s example is a refund window that moved from 30 days in March to 14 days in June, with the March rule dropping out.

Evaluation that scores the learning, not the vibe

You define what success means for your agent: was the problem solved, did the user have to correct it, did it end with a human. Reflexio scores conversations against that and traces every result back to the learnings that were retrieved. The API exposes session A/B tests, head-to-head response comparisons and retrieval holdout experiments. Each learning ends up with an impact verdict, so you can see which ones are doing the work and which ones are along for the ride.

A five-rung deployment ladder

Managed hosting gives each organisation an isolated schema. Bring your own keys covers OpenAI, Anthropic, DeepSeek, Qwen, xAI or a custom endpoint. You can keep the learning data in a Supabase project or Postgres instance you own while Reflexio runs the service. Bring your own cloud runs it inside your AWS, GCP or Azure account. Full self-hosting runs everything single-tenant with no connection back. The API your agent calls is the same on every rung, and a Governance API covers right-to-be-forgotten export and erase for enterprise accounts.

Does the Case Study Hold Up?

Reflexio published one customer story on 20 August 2026, about Creao, a platform that turns conversations into reusable agents. The numbers: scheduled playbook runs used a median of 2 model calls per delivered outcome against 8 for freeform chat tasks, a 75 percent difference, across a 28-day evaluation covering 465,305 runs. Playbook threads drew 36 percent fewer “Poor” verdicts from an automated quality judge across 230,477 evaluations, with the rates moving from 5.9 to 3.7 percent. Across 2,176 impact verdicts on retrieved learnings, 47.4 percent were positive, 51.4 percent neutral and 1.2 percent negative.

What earned our respect is the framing. The post calls the model-call comparison “an observed comparison across different task cohorts, not a causal estimate”, says the comparison “does not isolate how much of the difference was caused by the playbook itself”, and points out that playbook tasks are narrower than open-ended chat by design. The Product Hunt listing rounds this to “more than 30 percent” fewer failures and “more than 60 percent” token savings, which the one disclosed customer supports. It is still one customer, and the vendor did the measuring. But we rarely see a vendor volunteer the caveats before we go looking for them.

What Tripped Us Up

The package name will catch people. The open-source server installs with pip install reflexio-ai, currently at version 0.2.30, and there is a lighter reflexio-client at 0.2.16. The bare “reflexio” package on PyPI is an unrelated retry-policy library that was last updated in November 2025. Type the obvious thing and you get the wrong project. The docs are correct, but a search engine or an agent guessing the install command will not be.

It is Python first, and in practice Python only. The full documentation index has zero mentions of TypeScript. A Node agent talks to Reflexio over REST, which works, but there is no typed client and none of the lifecycle helpers. The naming is also inconsistent across surfaces: the pricing page sells Free, Pro and BYOC, while the docs split the world into “Hosted Enterprise” and “Local OSS”. Pro is Hosted Enterprise. It took us a few minutes to be sure of that, and a buyer should not need to be.

How Much Does Reflexio Cost?

Reflexio starts at $0 with a Free plan that includes 100,000 input tokens a month, 100 generated learnings, 1,000 search requests, data retained while the account is active, community support and a 30-day trial of Pro. Pro is $299 a month for 10 million input tokens, 10,000 learnings, 100,000 searches and priority support. BYOC self-hosted is custom, with no metered ceiling, configurable retention and dedicated onboarding. There is no annual toggle, and every Pro meter is exactly 100 times its Free counterpart.

Two things are missing. First, an overage rate: nothing on the pricing page or in the docs says what happens at 10,000,001 tokens, whether requests stop or a bill grows. Second, anything between $0 and $299. Mem0 goes free, then $19 a month for 50,000 add requests, then $249 for 500,000 add requests and 50,000 retrievals. Zep goes free with 10,000 credits, then $125 a month for 50,000 credits with $25 per extra 10,000, then $375. Letta starts at $20 a month. Reflexio’s free tier is generous enough to prove the loop works, and then the jump to Pro is the biggest in the category. The Apache licence is the escape hatch: self-host and the ceiling disappears.

Who is Reflexio Best For?

Use Reflexio if you run a production agent in Python that real users correct, and you want those corrections to change behaviour without a prompt-engineering sprint. It fits support, sales and coding agents in particular, and it fits regulated teams that need the learning data in their own database or their own cloud. If you already pay for mem0, the wrapper makes a side-by-side cheap.

Skip Reflexio if you are prototyping and only need conversational memory, if your stack is TypeScript first, or if you need a plan between free and $299. Skip it too if you want a managed service with a published overage rate you can model in a spreadsheet, because there is not one yet.

Best Reflexio Alternatives

Mem0 is the memory layer most people compare it with, and Reflexio ships a drop-in wrapper for it. Hobby is free, Starter is $19 a month with 50,000 add requests, and Pro is $249 a month with 500,000 add requests and 50,000 retrievals. It remembers; it does not evaluate whether remembering helped.

Zep pitches context engineering for agents. The free plan includes 10,000 credits a month with no rollover, Flex is $125 a month for 50,000 credits with $25 per extra 10,000, and Flex Plus is $375 a month for 200,000 credits. Credits roll over for 30 or 60 days on the paid plans.

Letta builds stateful agents rather than a learning layer you bolt on. There is a free plan capped at 3 stateful agents, a $20 a month plan, and an API plan billed at $0.10 per active agent per month plus $0.00015 per second of tool execution. Pick it if you want the runtime and the memory from one vendor.

Final Verdict: Is Reflexio Worth It?

Reflexio gets the hard part right. Learnings are scoped, inspectable and reversible, evaluation is built in rather than bolted on, and the deployment ladder runs from fully managed to air-gapped without changing the API. The open-source core is under a clean Apache 2.0 licence with a repository that moves daily, and the one case study reads like it was written by someone who expected to be checked.

What holds it at 7.5 out of 10 is commercial, not technical: a $0 to $299 cliff with nothing in between, no published overage, no annual pricing, and a Python-only SDK. Start on the free plan, run the 30-day Pro trial against a real agent, and measure the impact verdicts yourself. If they come back positive, the $299 will be easy to justify. If you cannot swallow the step, self-host it.