Syncing Memory Filesystem

MemFS Repository Repair

Use this skill only when the Git repository behind an agent's memory is not setting up or syncing correctly. For ordinary memory reads and edits, use the memory files without loading this skill. Delegate incidental upkeep during another task to the background memory subagent. When memory repair is the user's main request, diagnose and repair it directly, then verify the result before reporting success. If a memory worker you launched may still be running, wait for it: re-read its output file until it ends with [Task completed] or [Task failed] before editing memory files or changing Git state. While the harness reports a repair in progress, leave both alone as well; that worker edits $MEMORY_DIR in place.

Current Model

MemFS is a Git repository projected onto the computer where the agent is running. $MEMORY_DIR is the repository root. There is no second memory/ directory inside it.

The repository can use either memory layout. Inspect its current tree and the memory rules in the system prompt before editing files:

Root layout                         Existing layout
$MEMORY_DIR/                        $MEMORY_DIR/
├── MEMORY.md     # root index      ├── system/     # in-context memory
├── persona.md    # core memory     ├── reference/  # deferred memory
├── <topic>/                        └── skills/     # agent-owned skills
│   └── MEMORY.md # child index
└── skills/       # agent-owned skills

Cloud-backed agents have a hosted MemFS remote. Local-backend agents keep a local-only Git repository and do not need a remote or cloud credentials.

The agent doing the memory work commits its changes. After each turn, the harness pushes clean committed changes for cloud-backed agents. Local-backend commits remain on the current machine. Do not run git push for normal MemFS sync; let the harness push after the turn.

Committed memory changes do not alter the current compiled prompt immediately. Use /recompile when the current conversation must see changed core memory right away. Otherwise, the next prompt compilation or conversation will use the committed revision.

Start With the Harness

Prefer the harness commands over manual API calls, remote construction, or credential-helper edits:

/memfs status    # show whether MemFS is enabled and its path
/memfs enable    # initialize or repair MemFS setup
/memfs sync      # pull the hosted repository

From a shell, the standalone status and pull commands are:

letta memory status --agent "$AGENT_ID"
letta memory pull --agent "$AGENT_ID"

letta memory pull is a no-op for a local-backend agent because there is no hosted remote.

Do not reproduce /memfs enable by PATCHing agent tags or constructing a Git remote by hand. The enable flow also updates the system prompt mode, recompiles the agent, persists local settings, preserves and adds tags, initializes the checkout, installs hooks, configures identity, and seeds default memory files.

Inspect a Broken Checkout

Use $MEMORY_DIR instead of a hard-coded ~/.letta/agents/... path. Local and cloud-backed agents use different parent directories.

git -C "$MEMORY_DIR" status --short --branch
git -C "$MEMORY_DIR" remote get-url origin | sed -E 's#(https?://)[^/@]+@#\1<redacted>@#'
git -C "$MEMORY_DIR" log -5 --oneline

Do not print credential-helper values or tokens. Do not change global Git configuration. The harness installs or refreshes repository-local auth during clone and pull when the active transport supports a persistent helper. Desktop may instead use a temporary Git transport proxy and intentionally omit the persistent helper.

If the checkout is missing .git/, use /memfs enable. If it exists but is behind, use /memfs sync or letta memory pull --agent "$AGENT_ID". Pull also repairs recognized stale MemFS origin URLs and refreshes repository-local hooks, auth, branch tracking, and agent identity.

Uncommitted Changes

Raw file edits must preserve the active layout's rules:

• In the root layout, root and child MEMORY.md indexes have no frontmatter. Every other memory Markdown file has exactly name and description.
• In the existing layout, Markdown files under system/ and reference/ need a non-empty description. read_only is protected and cannot be added, removed, or changed by the agent.

---
description: What this memory file contains
---

Memory content goes here.

Review the complete diff before committing. Stage named memory files only and create a new commit. Once the repository is clean, the harness will push a cloud-backed agent's pending commits after the turn.