CogniCore Persistent, Transferable Experience for Claude Skills

[Kaushalt2004]

I’ve been working on CogniCore, an open-source memory layer for AI agents, and I wanted to share an experiment that I think could be interesting for Claude Skills.

Most memory systems focus on storing information or conversation history.

I’m exploring a different idea:

What if an agent could remember what it learned from actually completing a task and transfer that experience to another agent or session?

For example:

Claude Agent A
↓
Works on a task
↓
Approach A → Failed
Approach B → Failed
Approach C → Worked
↓
Tests / outcome verified
↓
CogniCore stores the experience
↓
Claude Agent B
↓
Similar task
↓
Relevant experience retrieved

Instead of giving Agent B the entire previous conversation, it can receive something like:

Previous experience:

Approach A failed because the module was stateful.

Approach B failed because the dependency version
was incompatible.

Approach C worked and was verified with the test suite.

Environment:
Python 3.12
Package version: X
Repository: Y

The goal is for the agent to understand what worked, what failed, and under what conditions.

Verification is important

I don't want an agent saying “this worked” to automatically become trusted memory.

I'm experimenting with:

Observed
↓
Evidence
↓
Verified
↓
Promoted
↓
Transferable
↓
Re-validated in new context

Experiences can also become outdated or be superseded by newer verified experiences.

Why Skills?

I think this could complement Skills rather than replace them.

A Skill provides the reusable instructions and workflow.

CogniCore could provide the experience accumulated from actually using that Skill.

So instead of:

"Here is how the agent should do this."

you could eventually have:

"Here is how the agent should do this, and here is what previous agents learned while doing it."

I've been testing this with coding-agent workflows, MCP, persistent memory, and experience retrieval.

CogniCore:
https://github.com/cognicore-dev/cognicore-my-openenv
I'd love feedback from people building Claude Skills:

Would transferable, verified experience be useful alongside Skills, or would you prefer Skills to remain completely deterministic/stateless?

I'm particularly interested in examples where you've seen an agent repeatedly make the same mistake or where knowledge learned by one agent/session would have helped another.