Without a shared research record
The conversation starts over.
You search old chats and notebooks, find the experiment, and explain the decision again. The conclusion may survive; its evidence and limits may not.
Local-first research memory
Research should not lose its memory.
Keep using Codex or Claude Code. RKA preserves the research decisions, evidence, and context that should survive your next session.
Core 3.0 availableNo RKA account required
The context you should not have to rebuild
You already tried that.
Keep the reason why.
Imagine a project predicting future events. Last week, you rejected a random data split. This week, a fresh AI session suggests it again.
With RKA connected to your agent
The reasoning is retrievable.
Ask your agent to retrieve the recorded decision and linked evidence. Review why it was made, where it applies, and whether it still holds before continuing.
Inspect the fictional decision and evidence
- Decision
- Use a time-based split for the next evaluation.
- Evidence
- The split-comparison experiment placed later observations in the training set under random splitting. That setup did not match the intended future-event evaluation.
- Scope
- This decision applies to this chronological prediction task. It is not a rule against random splitting in every study.
- Next check
- Revisit the decision if the evaluation goal or data collection process changes. Preserve the earlier rationale when recording a new decision.
This is an explanatory example, not a benchmark or a screenshot of a live RKA project. RKA makes recorded evidence retrievable; researchers still judge what it supports.
Your tools stay. The record carries forward.
A memory layer alongside your AI client.
Save what matters.
Use your connected agent or the local dashboard to record experiments, observations, decisions, and open questions.
Keep the evidence attached.
Link decisions and claims to their supporting records. Preserve uncertainty and the history of revisions, not just the latest summary.
Resume with context.
In your next session, ask the agent to retrieve the relevant RKA records. Inspect the sources and decide what to do next.
RKA Core stores and retrieves the record. Your chosen AI client reasons over what it retrieves; RKA Core is not another hosted chatbot.
Start with Core
Use the client you already know.
Install RKA on your machine, then connect your AI client through local MCP. You do not need App or Writer to use Core.
Setup instructions for macOS, Windows, and Linux. The documented quick start uses Docker Compose and a local MCP executable.
Open the complete installation guideStart RKA locally.
Follow the backend and MCP executable steps first. Keep the research database on storage you control.
Backend quick startChoose your setup instructions.
Both use local stdio MCP. Remote HTTP MCP and ChatGPT connectors are deferred in Core 3.0.
Ask: “List my RKA projects.”
An empty project list is expected on a fresh install. Then create a project and save one non-sensitive observation to try the workflow.
One record. Separate responsibilities.
Start with Core.
Follow what comes next.
Available · Core 3.0
RKA Core
The durable research record: journals, literature, decisions, claims, evidence, provenance, and retrieval. Works independently of the other projects.
Explore CoreIn development · Installation support
RKA App
Installation, runtime supervision, and optional deployment adapters around released Core artifacts. Not a required desktop interface.
Follow App developmentDesign phase · No supported authoring release
RKA Writer
A separate researcher-controlled writing workbench for connecting approved research meaning to a manuscript.
Explore the Writer designLocal-first, with explicit boundaries
Your research record.
Your control.
Local deployment is the recommended path for private or unpublished research. RKA Project does not operate a shared service that hosts everyone’s research records.
Read the access boundaryRecords stay with your installation.
In a local setup, RKA stores research records in a local database or Docker volume. You manage access, backups, and exports.
Your AI provider is a separate boundary.
Context retrieved into Codex or Claude Code may be sent to that client’s model provider. A remote embedding backend may also receive text. Choose clients and embedding settings that match your privacy needs.
Evidence remains inspectable.
Keep links to observations and decisions, including why they changed. Retrieved text is evidence to review, not an automatic guarantee that a claim is correct.
Status checked September 14, 2026
A usable core.
A lower barrier to entry.
Core maintenance, easier local setup, and Writer design have distinct milestones. A desktop app is not a prerequisite for getting started.
Core 3.0
The 3.0.0 release is available. Ongoing work focuses on correctness, recovery, retrieval, and compatibility.
Read the release notesMake local setup easier.
Improve agent-guided installation and released-image deployment through RKA App, without moving private research into a shared hosted service.
Try a sample. Own your instance.
A fixed-sample, read-only demo and a user-owned Hugging Face Space template are planned, not available from this page. They are optional trials, not a replacement for local deployment.
Keep authors in control.
Validate a separate writing workflow before offering a supported authoring runtime.
Read Writer’s current status