Integrate with anchordb

[kination]

Overview

Wire anchordb into devin as the persistent context store. Currently each session starts fresh — conversation history is held in memory and discarded on exit.

Motivation

apfel caps context at 4096 tokens. For longer sessions or multi-file refactors, relevant earlier turns get dropped. anchordb provides an append-only embedded store that can hold and retrieve past context without relying on the LLM's context window.

Proposed Work

• Add anchordb as a Cargo dependency (embed directly, replacing current subprocess approach)
• Persist each conversation turn (user + assistant) to anchordb on write
• On session start, retrieve recent relevant turns and inject them as synthetic history
• Add a --session <id> flag to resume a named session
• Add /history chat command to list stored turns

Notes

• anchordb integration was explicitly deferred during initial PoC phase
• Retrieval strategy (recency vs. semantic) TBD — start with recency