CogniCore

Agents don't share conversations. They share experience.

The Problem

Agent A discovers a bug fix, optimizes an algorithm, or learns what doesn't work. But when Agent B encounters a similar task, it starts from zero.

CogniCore solves this. It persists agent experiences across independent sessions, providing subsequent agents with retrieved context containing verified approaches—and failures to avoid.

See It In 30 Seconds

============================================================
 SESSION 1: Agent A (Claude) fixes an authentication bug
============================================================

AGENT A: Task -> Fix intermittent JWT authentication failures

Attempt 1: Increase JWT expiration
❌ FAILED: tests/auth timeout

Attempt 2: Modify retry logic
❌ FAILED: race condition in auth_middleware.py

Attempt 3: Fix refresh-token lifecycle
✅ SUCCESS
VERIFICATION: pytest tests/auth -> 18 passed

🧠 COGNICORE: Verified experience promoted and saved.

============================================================
 SESSION 2: Agent B (Gemini) encounters the same problem
============================================================

NEW SESSION: No previous conversation available.

AGENT B: Task -> Investigate JWT authentication failure.

AGENT B: Let me query CogniCore for related past experiences...

🧠 CogniCore retrieved 1 relevant experience
  ❌ Increasing JWT expiration
     Failed because: tests/auth timeout
  ❌ Modifying retry logic
     Failed because: race condition in auth_middleware.py
  ✅ Refresh-token lifecycle fix
     Verified: 18 tests passed

AGENT B:
"I found a verified previous experience for this repository.
 I'll inspect the refresh-token lifecycle first rather than
 repeating the two failed approaches."

Install

pip install cognicore-env

60-Second Quickstart

import cognicore

runtime = cognicore.CogniCoreRuntime()

def my_agent(task, context):
    print(f"Task: {task}")
    print(f"Memory: {context.get('experience')}")
    # Call your LLM here
    return True

result = runtime.execute(my_agent, task="Fix the JWT login bug")

Why This Isn't Just Memory

Most agent memory is just a dump of past conversations. CogniCore operates on Experience Memory.

Verification

An experience progresses through a strict lifecycle: Observed → Evidence → Verified → Promoted → Transferable

<<<<<<< Updated upstream memory = cognicore.Memory(max_size=10000) memory.store({"text": "add null check before dereferencing user", "category": "crash", "correct": True})

results = memory.semantic_search("null pointer crash", top_k=3)

=======
CogniCore requires independent evidence (e.g., `exit_code: 0` from tests) before allowing an experience to be transferred.

### Failure Memory
>>>>>>> Stashed changes

CogniCore explicitly stores what *didn't* work and why. Often, knowing a specific approach leads to a race condition is more valuable to an agent than a direct answer.

### Context Guardians

If an experience was verified on Python 3.11 but an agent retrieves it on Python 3.13, CogniCore flags a `CONTEXT MISMATCH` and requires the agent to independently verify the approach against the new environment before committing.

## Cross-Agent Transfer

CogniCore memory is model-agnostic. 

`Claude (learns) → CogniCore (verifies & stores) → Codex (retrieves & applies)`

---

## Benchmarks

**LongMemEval** (Strict R@5 on 500 chunks, 30 targets, 470 distractors)

| System         | Accuracy | Tokens / Query |
|----------------|----------|----------------|
| CogniCore FTS5 |  76.7%   |      68        |
| Mem0           |  70.0%   |      72        |
| Naive Context  |  95.0%   |    7,942       |

CogniCore achieves a ~99% token reduction vs Naive Context injection while outperforming remote embedding pipelines natively.

[Full methodology → docs/benchmarks.md](docs/benchmarks.md)

---

## Architecture & Features

CogniCore is a full cognitive framework for AI.

- [Architecture (Time Travel, Immutable Shield)](docs/architecture.md)
- [Semantic & Episodic Memory Details](docs/memory.md)
- [Experience Transfer Deep-Dive](docs/experience-transfer.md)
- [Integrations (MCP, API Keys, Claude Plugin)](docs/integrations.md)
- [Research (RL Environments, NEXUS Autonomous Coding)](docs/research.md)

---

## Community

🔥 Join the CogniCore Lab to experiment with agent memory.

- [Discord](https://discord.gg/cognicore)
- [Issues](https://github.com/cognicore-dev/cognicore-my-openenv/issues)
- [Contributions](CONTRIBUTING.md)