Roadmap

Roadmap

Implemented

• Claude Code and Codex capture.
• Normalized event schema.
• Markdown transcript rendering.
• Signals and aggregate reporting.
• Mining methods and risk-aware filtering.
• SQLite memory schema and rebuildable index.
• FTS5 keyword retrieval with scope filtering.
• [[wiki-link]] graph extraction and one-hop expansion.
• Authored markdown import.
• Utility writes and q-value updates.
• Prompt-time memory weave.
• Static dashboard with signals, cost, mined memory, and indexed memory sections.

Next

• Optional embedding tier with sqlite-vec and pure-Python fallback.
• Deeper reward wiring from trajectory signals into memory utility.
• Promotion workflow for candidate to accepted memories.
• Export reviewed memories into Codex-compatible skills.
• Progressive disclosure for exported skills.
• More dashboard sections for deprecated memories and skills needing review.

Non-Goals For V1

• No Postgres, pgvector, Neo4j, or server requirement.
• No FAISS, LangChain, or hard numpy dependency.
• No automatic live injection into Codex or Claude sessions before retrieval risk controls mature.