This document captures the decisions that define the public 1.0.x release
line.
The durable store is local SQLite. This keeps memory inspectable, portable, and easy to reason about across MCP-compatible AI agents and coding workflows.
Runtime adapters are intentionally thin. Codex notify is one capture adapter;
MCP retrieval is the surface shared by Codex, Claude Code, Google Gemini CLI,
Qwen Code/Ollama workflows, and other local MCP clients.
The server is not a single broad menu for every client. Tool exposure is controlled by profile:
minimal for low-impact continuity checks;standard for normal read-oriented retrieval and diagnostics;full for the complete local surface.--read-only remains the safety boundary for agent workflows that should not
write snapshots, policies, inheritance, repairs, or closure records.
structuredContent is the canonical MCP result. content.text is only the
client-facing text view, controlled by --response-mode. List results use an
object root shaped as {items, count} for stricter MCP clients.
The project stores turns and extracted observations first. More advanced semantic enrichment can come later, but the baseline must remain deterministic, auditable, and reproducible with local tests.
Useful memory is not only a summary of past decisions. The continuity pack also tracks objectives, open work, blockers, DoD gaps, recent changes, and scope guardrails so agents do not falsely declare completion.
pack_hash, known_pack_hash, compact packs, and response modes reduce repeated
operational context when there is repeated context to reuse. Public percentage
claims must stay tied to reproducible fixtures or documented live evidence.
The public core does not claim hosted bridges, OAuth, ChatGPT web, Claude web,
embeddings, vector retrieval, or built-in database encryption. Those areas stay
outside the 1.0.x core unless a future release explicitly changes scope.