Skip to main content
Changelog
Live

The knowledge base is now organized by memory layer

By CodeLantern Team ·

The .codelantern directory tree beside the five agent memory layers its files map to: short-term scratchpad, long-term semantic, long-term procedural, long-term episodic, and reasoning..codelantern/.codelantern/knowledge-base/work-items/item-5/item-7/FILE → MEMORY LAYERShort-termTHE SCRATCHPADAgent context window — we don't persist itLong-term semanticWHAT IS TRUEThe project's shape, its rules, and hard-won lessonsLong-term proceduralHOW TO DO ITThe steps for the things you do oftenLong-term episodicWHAT HAPPENEDThe record of past work, and how it wentReasoningWHY WE CHOSEThe decisions, and that they are settled

A single markdown file has become a memory system, with a different layer for each kind of thing an agent needs to know.

The CodeLantern knowledge base is now a set of files that give an agent a proper memory system, each layer holding a different kind of information for a different purpose: semantic for what is true about the project, procedural for how things are done, episodic for what happened, and reasoning for why a choice was made and that it’s settled.

Why it’s shaped like human memory

Psychologists stopped treating memory as one thing decades ago, separating knowing that something is true from remembering something that happened, and setting knowing how to do a thing apart from both. We’ve applied that learning to how we context engineer agents: an agent, like a person, meets more than it can hold. Giving it the structure the developer steering it already carries is how you bring it up to speed on a codebase, much the way you bring on a junior developer.

In practice, skills read less to know more, and each session starts from what the last one learned rather than from scratch.

Lives in your repository, synced with the platform portal

It is all markdown in your repository. You own it, you review changes to it the way you review anything else, and you can edit it by hand. What gets captured syncs to CodeLantern, where entries become searchable for later sessions and organization admins can see the knowledge compound.

The portal's Knowledge tab: 1,230 knowledge base entries captured across 11 repositories, 677 this month; a type breakdown running from 37% learnings and 31% conventions down to 3% project facts; and a table of recent entries with their type and age.Analytics · KnowledgeTOTAL ENTRIES1,230REPOS WITH A KB11ENTRIES THIS MONTH677TYPE BREAKDOWNLearnings 37%Conventions 31%Glossary 11%Recipes 7%Components 7%ADR 4%Project 3%TYPEENTRYAGERecipesRecipes > Add a database migration2dLearningsDO > Seed the test database before the suite runs2dConventionsConventions > Name feature flags with the team prefix3dADR0014. Keep the public API surface versionless4dGlossaryGlossary > Provisioning code4dComponentsComponents > DataTable5d
The Knowledge tab: what each repository has captured, broken down by type, with the entries themselves underneath.

Availability

Live for onboarded CodeLantern customers with configured repositories and a connected GitHub App. Knowledge analytics in the portal are available to organization admins.