v3.7.0 release • Track D memory reliability for Codex

Persistent memory for AI coding tools

Locus gives Codex and Claude Code a local memory layer for project structure, durable decisions, searchable history, summary-first memory_recall, and diagnostics. Built on MCP. Safe-by-default. Designed so your agent stops starting from zero.

Русский 简体中文 Codex CLI Codex desktop / extension Claude Code MCP server
install $ npx -y locus-memory@latest install codex
One command installs the Codex skill and configures a version-pinned MCP runtime.
codex + locus memory
~/project $ codex
Loading project context…
Locus MCP connected
memory_search("auth schema decisions")
Locus: found prior auth trade-offs and recent Codex import status
Codex: continuing from the JWT vs sessions decision we made earlier
You: finish the auth module and keep the previous token model

How Locus compares

Locus is intentionally not a full agent runtime or cloud memory platform. It is a smaller local MCP memory layer for coding agents, with Codex as the first-class product path.

Project
Main strength
Why Locus is different
agentmemory
Broad coding-agent memory stack with many tools, hooks, viewer, and benchmark claims.
Locus is smaller, simpler, Codex-first, and ships as one npm MCP runtime.
AIDE Memory
Path-scoped local memory and tiny context nudges.
Locus focuses on MCP tools, Codex JSONL import, diagnostics, and explicit recall UX.
Mem0 / Letta / Zep
Larger agent memory platforms, SDKs, stateful runtimes, or graph infrastructure.
Locus plugs into existing coding tools instead of asking users to run a bigger platform.

What ships today

Locus is already a real product line for Codex, not a future promise. The current release focuses on Track C richer recall, diagnostics, and one-command install.

Codex memory that survives the next session

Track C validates redacted Codex recall through `memory_recall`: dated Russian questions, capture decisions, rejected alternatives, user style, npm errors, next steps, and validation facts pass through import, durable memory, and recall.

Safe-by-default capture

`redacted` is the recommended rich recall mode for Codex. `metadata` remains diagnostics-first. `full` is maximum recall and requires an explicit privacy warning; never treat it as risk-free.

Diagnosable, not magical

`memory_status` and `memory_doctor` expose Codex-specific state so users can see what is imported, what is missing, and where the failure actually lives. `candidateGroups` make ambiguous recall inspectable instead of magical.

One-command Codex install

Track B packages Locus as `locus-memory`, adds installer, doctor, uninstall commands, and generates a marketplace bundle while keeping manual MCP fallback supported.

Why teams keep losing context

Most coding agents are strong inside a session and weak across sessions. Locus exists to bridge that gap without turning your workflow into a ritual of manual summaries.

Project structure

Scan the codebase, map files and exports, and keep that structure searchable through MCP.

Decision memory

Save architectural choices and trade-offs so the next session starts with context, not guesswork.

Conversation history

Carbon Copy-style event ingest captures what happened, then makes it searchable through the same memory layer.

Operational trust

Diagnostics, audit visibility, and explicit capture modes matter as much as recall itself.

Explore the real project

This landing page is intentionally simple. The real source of truth remains the repository: Quick Start, Codex setup, roadmap, releases, and implementation details all live there.

Copied