Continual Memory Policy Model — AI skill for Claude Code

Continual Memory Policy Model

[](#experimental-status-and-safety) [](pyproject.toml) [](LICENSE)

[!CAUTION] **Experimental research software.** This repository is not a production memory service, medical or legal record system, or autonomous online-learning system. The included evaluations use small, partly authored trajectories. Keep human review, privacy controls, evaluation gates, and rollback in the loop.

An executable research scaffold for a small model that decides **what an AI agent should remember**. Facts stay in an external SQLite memory store. The policy chooses `WRITE`, `UPDATE`, `DELETE`, `LINK`, `COMPACT`, or `NOOP`, then learns periodically from what happened after those choices.


The idea in one minute

The project separates two kinds of learning:

1. Memory changes immediately. A conversation can add, revise, connect, combine, retire, or ignore information in the external database.
2. Model weights change later. Outcomes are collected into a training buffer. A candidate LoRA adapter is trained, tested, and promoted only after an explicit human decision.

That means a useful memory can be written at step 40, recalled at step 8,000, and receive delayed credit when the later task succeeds. A harmful or stale memory can receive negative credit. Credit is an inspectable heuristic, not proof that the memory caused the result.

What people see

The **Memory Center** is a local, read-only UI for everyday inspection. It uses plain language first and keeps the technical evidence one level deeper.


The UI includes:

• Home: what is