Carry context into later work

Localix keeps a memory folder per repository, agent definitions, reusable skills, and searchable session transcripts for later work. Use them to preserve durable facts, repeat procedures, or retrieve evidence from an earlier session.

Boundary: Localix loads stored context, but it does not extract every durable fact or workflow automatically.

What you can do

How it works

Memory

Memory is Claude Code's auto memory, in Claude Code's own folder, so a repository has the same memory in both apps:

~/.claude/projects/<project>/memory/
├── MEMORY.md          # the index: one line per memory
├── user-role.md       # one fact per file
└── deploy-days.md

<project> is the repository's main checkout with every character but letters and digits turned into -, so every worktree and subfolder of a repository shares one folder; outside a repository it is the working folder. CLAUDE_CONFIG_DIR moves the whole tree, and autoMemoryDirectory in your own Claude Code settings names another folder (a repository's committed settings cannot). autoMemoryEnabled: false or CLAUDE_CODE_DISABLE_AUTO_MEMORY=1 turns memory off.

The agent saves a memory with the ordinary file tools: one fact per file, starting with frontmatter (name, description, and type: user | feedback | project | reference under metadata), then a one-line pointer in MEMORY.md. Writes into the run's own memory folder raise no permission card; another project's folder still asks. Topic files are not loaded on their own: the agent reads one with Read when the index says it matters.

Prompt snapshots

Each session has a stored prompt snapshot for its context boundary. MEMORY.md and project instructions normally stay frozen across later runs, including API restarts, as Claude Code loads MEMORY.md once per session. A context checkpoint or a new session creates a fresh boundary; so does a run in another repository, whose memory folder differs. Non-chat runs build fresh context rather than using the chat snapshot. The bounded skill catalog also refreshes when its discovered content or overrides change.

A memory saved mid-session is on disk at once, and the agent can Read it, while the frozen prompt keeps the index it started with. Set LOCALIX_FREEZE_CONTEXT=0 to rebuild the snapshot on each run.

Recall

GlobalSearch recalls past work. A query runs the search palette's own search — words, and meaning when semantic search is on — with domains narrowing it to categories such as messages or chats; around_message_id or session_id reads a bounded window of a conversation; no arguments browse recent sessions. The tool returns previews and message identifiers rather than replaying a full transcript.

An external memory provider can run loadContext when a fresh chat prompt snapshot is built, for a run that keeps memory. It may append to the MEMORY.md index or replace it in the prompt; the folder and its files stay. The 8,000-code-unit cap applies to provider output; append mode can produce a larger combined context with the index. Its onRunEnd hook receives each completed turn's user and assistant text.

Configuration

Limits

• A model may decide not to save a memory or a skill. Use /memory <fact>, agent instructions, or an explicit skill save when persistence is required.
• Memory folders are shared with Claude Code: what either app saves, the other loads.
• Memory Markdown is plaintext on disk. It is readable by processes with access to the same OS account and home directory.
• GlobalSearch must survive the active role and tool policy. Discovery also requires SQLite FTS5; an unavailable FTS table returns an error instead of falling back to transcript injection.
• A whole-tool denial without later argument exceptions removes GlobalSearch or Skill from the roster. Conditional rules can retain a tool while refusing individual calls. A run without Write and Edit keeps no memory.
• A memory provider's onRunEnd hook is fire-and-forget. Failure or timeout does not roll back the completed chat turn. A hook timeout stops waiting; its work can continue because the hook port has no abort signal.

Related guides

• Skills
• Sessions and projects
• Context checkpoints
• Mini apps

Source of truth

• Memory folder, index loading and /memory <fact>
• Memory section and the folders a run writes without a card
• Persisted prompt snapshot and refresh exceptions
• Agent prompt and provider admission
• Prompt loading and snapshot cache
• History recall
• /memory command
• Skill tool and skill writes
• Completion hook invocation
• Memory provider lifecycle