GitHub - kurikomi-labs/komi-learn: Continuous memory + self-improvement for AI agents. Learns how you work, recalls it automatically, no commands. Claude Code & Codex. · GitHub

Skip to content

Navigation Menu

Toggle navigation

Sign in

Appearance settings

PlatformAI CODE CREATIONGitHub CopilotWrite better code with AIGitHub SparkBuild and deploy intelligent appsGitHub ModelsManage and compare promptsMCP RegistryNewIntegrate external toolsDEVELOPER WORKFLOWSActionsAutomate any workflowCodespacesInstant dev environmentsIssuesPlan and track workCode ReviewManage code changesAPPLICATION SECURITYGitHub Advanced SecurityFind and fix vulnerabilitiesCode securitySecure your code as you buildSecret protectionStop leaks before they startEXPLOREWhy GitHubDocumentationBlogChangelogMarketplaceView all featuresSolutionsBY COMPANY SIZEEnterprisesSmall and medium teamsStartupsNonprofitsBY USE CASEApp ModernizationDevSecOpsDevOpsCI/CDView all use casesBY INDUSTRYHealthcareFinancial servicesManufacturingGovernmentView all industriesView all solutionsResourcesEXPLORE BY TOPICAISoftware DevelopmentDevOpsSecurityView all topicsEXPLORE BY TYPECustomer storiesEvents & webinarsEbooks & reportsBusiness insightsGitHub SkillsSUPPORT & SERVICESDocumentationCustomer supportCommunity forumTrust centerPartnersView all resourcesOpen SourceCOMMUNITYGitHub SponsorsFund open source developersPROGRAMSSecurity LabMaintainer CommunityAcceleratorGitHub StarsArchive ProgramREPOSITORIESTopicsTrendingCollectionsEnterpriseENTERPRISE SOLUTIONSEnterprise platformAI-powered developer platformAVAILABLE ADD-ONSGitHub Advanced SecurityEnterprise-grade security featuresCopilot for BusinessEnterprise-grade AI featuresPremium SupportEnterprise-grade 24/7 supportPricing

Search or jump to...

Search code, repositories, users, issues, pull requests...

Search

Clear

Search syntax tips

Provide feedback

We read every piece of feedback, and take your input very seriously.

Include my email address so I can be contacted

Cancel

Submit feedback

Saved searches

Use saved searches to filter your results more quickly

Name

Query

To see all available qualifiers, see our documentation.

Cancel

Create saved search

Sign in

Sign up

Appearance settings

Resetting focus

You signed in with another tab or window. Reload to refresh your session.
You signed out in another tab or window. Reload to refresh your session.
You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

kurikomi-labs

/

komi-learn

Public

Notifications
You must be signed in to change notification settings

Fork
0

Star
6

Code

Issues
0

Pull requests
0

Actions

Security and quality
0

Insights

Additional navigation options

Code

Issues

Pull requests

Actions

Security and quality

Insights

kurikomi-labs/komi-learn

 mainBranchesTagsGo to fileCodeOpen more actions menuFolders and filesNameNameLast commit messageLast commit dateLatest commit History42 Commits42 Commits.claude-plugin.claude-plugin  .github/workflows.github/workflows  examplesexamples  hookshooks  komikomi  pool-repo-templatepool-repo-template  teststests  .gitignore.gitignore  LICENSELICENSE  README.mdREADME.md  pyproject.tomlpyproject.toml  View all filesRepository files navigationREADMEMIT licensekomi-learn
Continuous memory and self-improvement for coding agents. It learns how you work and recalls it automatically, with no commands. Works with Claude Code and Codex.

It watches a session, distills durable lessons in the background (your style, your stack, fixes that worked), and loads the relevant ones at the start of the next session. No slash commands, nothing to save by hand.
The idea is from Hermes Agent; this is my own take, generalized across hosts with an optional shared layer (the community pool, below).
It's early. The core loop is built and CI-tested, but it hasn't been battle-tested across a lot of real sessions yet. Feedback and bug reports are welcome.
Install
pip install komi-learn
komi-learn install # or: komi-learn install --host codex
install runs a short interactive setup, then recall and background learning start in your next session. If you already use Claude Code you're already logged in. For scripts, komi-learn install --yes takes the defaults.
From source:
git clone https://github.com/kurikomi-labs/komi-learn
cd komi-learn
pip install -e .
Commands
komi-learn doctor # check the install and what to fix
komi-learn status # config + how much it has learned
komi-learn config # change any setting (menu, or `config set <key> <val>`)
komi-learn sync # pull the latest community learnings
komi-learn queue # review/approve/reject what you'd contribute to the pool
komi-learn forget <x> # erase learnings matching <x> (archive, or --hard to delete)
komi-learn uninstall # remove the hooks (keeps your data; --purge to wipe)
You can change anything after install, e.g. komi-learn config set recall.semantic false or leave the pool with komi-learn config set pool.repo_url "".
How it works

Recall: at session start, learnings relevant to the current context are loaded.
Distill: after the session, a background pass reads the transcript and extracts durable lessons (corrections, techniques, fixes).
Curate: over time it merges overlapping lessons and archives stale ones.
Share (optional): general lessons can be contributed to the community pool, but only ones you approve.

It tries not to learn the wrong things. Secrets, machine-specific paths, one-off failures, and "tool X is broken" complaints are filtered out by a deterministic check before the LLM ever sees them.
Community pool (optional)
A public pool of general agent lessons, stored as a GitHub repo of signed Markdown files (no server). If you opt in, you get lessons other people's agents figured out, and you can contribute your own.
Contributions are scrubbed of anything identifying and never leave your machine without your approval (each one opens a PR you reviewed). Learnings are content-addressed (BLAKE3) and signed (Ed25519); one signed by more distinct GitHub accounts ranks higher when pulled. That account count is Sybil-resistant but not Sybil-proof, so it's an advisory signal, not a hard trust gate. Recalled community items are labelled and treated as untrusted input. Details: pool-repo-template/CONTRIBUTING.md.
Try it offline
No setup or API key needed:
python examples/demo_loop.py
It runs two sessions: you correct the agent in the first, and the second shows it recalling that with nothing typed.
Requirements

Python 3.10+
Claude Code or Codex (the agent it plugs into)
A working model for the distill step: your existing Claude Code login, or komi-learn login, or an API key via --api-key.

komi-learn install verifies these with a real model call and stops with fix steps if something's missing. At runtime, if a hook can't reach the model it skips that learning pass rather than interrupting your session.
The engine has no required dependencies. Optional extras add real signing (pip install komi-learn[crypto]) and local semantic recall ([smart]); without them it falls back to a stdlib hash and keyword search.
To run your own pool, see pool-repo-template/.
MIT. Issues and PRs welcome.

About

Continuous memory + self-improvement for AI agents. Learns how you work, recalls it automatically, no commands. Claude Code & Codex.

Topics

codex

ai-agents

continuous-learning

llm

claude-code

Resources

Readme

License

MIT license

Uh oh!

There was an error while loading. Please reload this page.

Activity

Custom properties
Stars

6
stars
Watchers

0
watching
Forks

0
forks

Report repository

Contributors

Uh oh!

There was an error while loading. Please reload this page.

Languages

Python
100.0%

Footer

© 2026 GitHub, Inc.

Footer navigation

Terms

Privacy

Security

Status

Community

Docs

Contact

Manage cookies

Do not share my personal information

You can’t perform that action at this time.

The komi-learn project is designed to implement continuous memory and self-improvement capabilities for AI agents by enabling them to automatically learn from user workflows and recall that knowledge without requiring explicit commands. This approach draws inspiration from the concept of the Hermes Agent, focusing on an architecture where the agent can internalize and apply learned patterns. The core mechanism involves several coordinated processes: recall, distillation, curation, and optional sharing within a community pool.

The recall process dictates that upon the start of any session, the system loads the learnings most relevant to the current context. Following the session, a background distillation step processes the session transcript to extract durable lessons, which include corrections, effective techniques, and successful fixes. These lessons are then curated by merging overlapping insights and archiving any stale information. A crucial safety feature is implemented to ensure quality: the system filters out highly specific or transient failures, such as secrets, machine-specific paths, and one-off errors, before these inputs are processed by the large language model, thereby preventing the agent from learning erroneous or insecure information.

The system optionally integrates a community pool for sharing general agent lessons. If enabled, lessons can be contributed by users, but these contributions are subject to strict vetting and safety protocols. Contributions are content-addressed using BLAKE3 and signed using Ed25519 cryptography, which means the trust signal is derived from distinct GitHub accounts; this mechanism provides an advisory signal regarding trust rather than a hard trust gate, ensuring that contributions remain on the user's machine without explicit approval.

Installation and operation require Python 3.10+ and access to an underlying code model such as Claude Code or Codex. The installation process includes setup steps that verify dependencies, and runtime behavior is managed through distinct commands that allow users to manage configuration, check status, synchronize with the community pool, review contributions, and explicitly forget specific learnings. The engine is modular, allowing for optional extras such as cryptographic signing or semantic recall, which fall back to standard hashing if those modules are not implemented. The system is built upon a framework where the agent can interact with the underlying code generation models while maintaining an internal, evolving memory of how successfully those interactions have gone.