dsh-anagenesis: A Rollback-Capable Agent Memory Orchestration Layer for DeepSeek Harness

What Does It Do?

Anagenesis is a DSH plugin for DeepSeek Harness that turns agent memory into a programmable orchestration layer plus a self-evolution engine. It solves the problem of black-box automatic memory, where the background decides what to store and how to recall it: here the agent explicitly calls memory tools, injection strategies can be swapped at runtime, and every change ships with an inverse operation that can be rolled back. Storage is self-contained, offline, and has zero runtime dependencies, with a terminal dashboard and text diagrams that make the memory store visible.

Core Features

• Intent-driven recall: retrieves memories by intent such as orient, recall_precedent, avoid_mistake, and reuse_procedure, each carrying its own kinds, states, confidence floor, weights, and token budget instead of keyword or vector similarity.
• Explicit memory lifecycle: write, promote, demote, lock, expire, and forget operations move entries through draft → active → verified → locked, with deletion leaving only an auditable tombstone.
• Swappable injection strategy stacks: named stacks including explore, exploit, debug, distill, recon, and crisis are stored per scope, so one agent switching stacks does not affect another, while guard stays at the bottom as an invariant.
• Reversible changes: every transaction returns an inverse function, and inverse operations are written to the journal alongside the forward patch, so rollbacks still work after a restart; strategy switches and tuning rollbacks build on this primitive.
• Bounded self-tuning: the meta layer adjusts confidence floors and token budgets from feedback under a parameter envelope with UCB1 arm selection, audit, evaluation, and rollback; only tier-1 is self-modifiable while the outer ring stays frozen.
• Visibility and guardrails: a terminal dashboard, Mermaid/D2/ASCII diagrams, and an optional desktop window expose the store, backed by monotonically tightening invariants, circuit breakers, and safeMode.

How to Use This Plugin?

After enabling it in DSH and restarting the profile, pick the agent preset named anagenesis when creating a new session; the default strategy stack is guard + exploit. Then just work normally: the agent calls recall and write tools on its own, so you never write a query language. To inspect the store, have it call the dashboard tool for a one-frame TUI overview, or the diagram tool for relationship, strategy-timeline, and lifecycle-flow charts you can paste into Markdown. To adjust behavior, use the governance tools to switch stacks, tune and roll back the meta layer, or review status, journal, audit, and health; you can also have the agent open the desktop window tool to watch memory changes in a separate window.

How to Troubleshoot This Plugin?

When the memory store misbehaves, first use the governance tools to review status, journal, audit, and health to identify which change introduced the problem. If a strategy switch or tuning step caused drift, roll back that operation; inverse operations are written to the journal alongside the forward patch, so they remain valid after a restart. If the whole system looks unstable, enable safeMode to freeze self-modification and then work through the issue per the official instructions. Once the journal retention policy has dropped content, the status view reports it explicitly rather than pretending that sequence number never existed.