# Core Memory (Full Context)

> Durable, local-first memory runtime for agents that need grounded recall of why a decision happened, not only what happened.

## What Core Memory solves

Most agent memory stacks can retrieve semantically similar text. Core Memory focuses on causal continuity:

- Capture finalized turns with typed memory events.
- Preserve causal links and chronology across sessions.
- Keep retrieval grounded to source turns and artifacts.
- Support hydration so results can point back to evidence.

## Key architectural novelties

1. Deterministic causal recall:
   - Explicit tie-break policy.
   - Stable ordering across restarts.
   - Explain reports can include bead ids and replay hashes.

2. Two-factor promotion with adaptive threshold:
   - Promotion requires score + reinforcement.
   - Threshold self-raises as promoted ratio grows.
   - Prevents promotion inflation.

3. Session-start rolling-window injection:
   - Token-budgeted continuity window from distilled historical beads.
   - Deterministically selected and idempotent.
   - Preserves continuity at every new session start.

## Install

```bash
pip install core-memory
```

## Main links

- Website: https://www.usecorememory.com/
- Repository: https://github.com/JohnnyFiv3r/Core-Memory
- README: https://github.com/JohnnyFiv3r/Core-Memory#readme