Self Configuration

Self-Configuration

Use this skill when the user asks you to change yourself or the Letta Code runtime around you.

The important part is choosing the right layer. Do not smear a preference into deterministic config, and do not bury a deterministic safety rule in prose memory.

First choose the layer

Decision rule: if the model should remember and reason about it, use memory. If the runtime must enforce it or route it before the model decides anything, use settings, API fields, mods, channels, or schedules.

Safe workflow

1. Identify scope: current conversation, current agent, project, or global user config. Model changes target the current conversation unless the user asks about the agent default.
2. Inspect current state first and save the relevant safe fields as a rollback patch. Do not copy secrets or full compiled prompts into backups.
3. Prefer a dry run for API patches and scripts.
4. Apply the smallest change that satisfies the request.
5. Verify the effective state after the write.
6. Tell the user what changed and whether a restart/new conversation is needed.

Never print secrets. If inspecting env settings, list keys unless the user explicitly asks for values and the values are safe to reveal.

Guardrails are not security boundaries

These helper scripts reduce accidental harm. They are not a security boundary against an agent with unrestricted Bash, raw curl/SDK access, API credentials, or filesystem access. LETTA_API_KEY and the installed CLI may have authority over other agents visible to the same account/server.

Never target another agent or conversation unless explicitly directed and verified. If AGENT_ID or CONVERSATION_ID is set, the server-setting helpers reject mismatched live/GET operations unless --allow-other-agent is present. If the current env ID is absent, explicit IDs remain usable for out-of-band recovery.

If a broken model or prompt prevents the agent from completing a turn, recover out of band from another shell or client with the CLI/API. Do not depend on the broken model to repair itself.

Inspect effective state before changing it

Local settings, server state, and the current process are different sources of truth. Inspect the layer you intend to change before writing it.

• letta model list [--byok | --hosted] lists available models.
• letta model set [model_handle] [--reasoning <reasoning-option>] [--model-settings <json>] [--default] changes the current conversation's model, reasoning, or individual model_settings fields; add --default only when the user asks for the agent default.
• letta model get [--default] gets the current model configuration, including the full redacted model_settings; --default gets the agent's default configuration.

letta model reads and writes through the active backend, so it works the same on every backend. Prefer it over the REST helper scripts for anything model-related.

Account credits and model quota

Run letta usage for a Markdown overview of the current plan, credit balance, and letta/* model quota (lettaTier only). Report the server's bucket (full, high, medium, low, or empty) and quota/daily reset timestamps as-is; do not infer exact requests or percentages. Amounts are credits, not dollars; preserve negative balances. An omitted daily reset is shown as unavailable.

The command uses CLI auth and respects LETTA_API_KEY/LETTA_BASE_URL, not agent or conversation selectors. Credits belong to the organization; user-scoped quota belongs to the authenticated user, not necessarily the person chatting with the agent. In local mode, use letta --backend cloud usage only when the user wants Cloud account usage.

Use letta model list for available models; credits and quota buckets do not guarantee inference availability. letta usage does not include session token statistics; the interactive /usage command is a separate surface. If either lookup fails, the command exits nonzero without partial usage. Treat that as unavailable data, not zero credits or exhausted quota.

Billing path when changing models

The same model can often be reached through more than one route: a connected subscription (for example a ChatGPT or Grok plan), the Letta plan (letta/*), or per-token billing against organization credits or the user's own API key. Users choose provider names, so a handle's prefix does not reliably show which route it bills through.

Before switching models, consider how the current model is billed and keep the user on that route unless they asked to change it. Use the current handle, the labels in letta model list, and anything the user has said about billing as evidence. If several available handles serve the requested model and you cannot tell which one uses the user's subscription, list the candidates and ask before switching. Do not silently move a user from a subscription to per-token billing.

letta model list --byok includes both connected subscriptions and user API keys, so it does not separate the two. letta usage covers only Letta credits and letta/* quota, not connected subscriptions.

Harness and server settings

Use the secret-safe local/runtime report for harness settings, permissions, and backend diagnostics:

python3 <SKILL_DIR>/scripts/show_config.py --cwd "$PWD"

Read each model scope with letta model get before changing it:

letta model get                               # current conversation (or agent if none)
letta model get --default                     # agent default
letta model get --conversation "$CONVERSATION_ID"

Do not infer an agent default from one conversation or infer a conversation override from the agent. Report both when diagnosing model or context differences.

If CLI behavior does not match the docs, stop and inspect command -v letta, type -a letta, and letta --version. A stale or shadowed binary is a config bug, not a reason to guess.

Memory and identity

Use memory when the user wants you to remember, prefer, learn, or change your identity/personality.

Inspect the projected memory tree in the system prompt before choosing paths. Letta Code supports two layouts:

Use the active layout shown by the prompt and memory files. Do not create a system/ directory in a root-layout repository or move existing-layout memory to the root as part of an unrelated self-configuration request.

A requested memory or identity change is memory as the main task: make the edit directly with ordinary file tools, preserve the active layout, stage only the files you changed by explicit path, commit in the same shell call, and verify the result before reporting success. If a memory worker you launched may still be running, read its output file first (it ends with [Task completed] or [Task failed]) and reread the files before editing what it was asked to change. Delegate to the background memory subagent only when the change is incidental to another task the user is waiting on.

Do not use API system-prompt replacement for ordinary learning. That can clobber the compiled prompt. Edit the memory files instead.

Server-side agent and conversation settings

Server fields control model execution and agent metadata. Change the current conversation for model changes. Change the agent default only when the user asks for it.

Model and model settings

Use letta model for the model handle, reasoning level, and any model_settings field (for example the output-token limit, temperature, or parallel_tool_calls). It goes through the active backend, so it needs no extra credentials or base URL.

letta model set "openai/gpt-5.2"                          # select a model (applies its defaults)
letta model set --reasoning high                          # change reasoning only
letta model set --model-settings '{"max_tokens":16384}'          # use the key `letta model get` shows
letta model set --model-settings '{"temperature":0.2}' --default   # agent default
letta model get                                           # verify

--model-settings takes a JSON object and shallow-merges it into the target's current model_settings (the conversation override when a conversation is targeted, otherwise the agent default). It keeps the model and every field you do not name. Keys are written as given, so use the field names that letta model get shows for the target; for example the output-token limit is stored as max_tokens on some backends and max_output_tokens on others. Nested objects such as reasoning are replaced as a whole, so include every nested key you want to keep. With a handle or --reasoning, the model selection is applied first and the given fields are merged on top. Selecting a new handle resets settings to that model's defaults, so reapply custom fields in the same command if needed.

Provider reasoning fields differ. Read references/model-settings.md before changing reasoning or provider-specific settings.

REST helper for other fields

scripts/update-agent-settings.ts covers fields letta model does not: context_window_limit, name, description, and system prompt replacement. It talks to the REST API at LETTA_BASE_URL and requires LETTA_API_KEY, so it only works against a server that exposes the REST API; it does not work on the embedded local backend.

Required environment for live API writes:

export LETTA_API_KEY=...
export AGENT_ID=agent-...
export CONVERSATION_ID=conv-...   # only needed for conversation-scoped changes
export LETTA_BASE_URL=...         # required; use the current server, not a hard-coded Cloud URL

The scripts in this skill default to AGENT_ID, CONVERSATION_ID, and LETTA_BASE_URL. Server reads and writes require LETTA_BASE_URL or explicit --base-url; they never silently fall back to api.letta.com. Keep LETTA_BASE_URL paired with the LETTA_API_KEY supplied by the current runtime so local, self-hosted, and non-default Cloud environments are not accidentally redirected. Pass explicit IDs when there is any doubt. --show fetches the selected agent or conversation and prints only safe effective fields. Server operations reject target IDs that differ from the current env ID unless --allow-other-agent is passed. Dry-run output is labeled: offline_partial_patch means no server state was fetched; effective_merged_patch means the script fetched current server state and shows the merged patch that would be sent.

Dry-runable update script

npx tsx <SKILL_DIR>/scripts/update-agent-settings.ts --help

Patch the current conversation's context window limit:

npx tsx <SKILL_DIR>/scripts/update-agent-settings.ts \
  --target conversation \
  --conversation-id "$CONVERSATION_ID" \
  --context-window-limit 64000 \
  --dry-run

Patch the agent default only when the user asks for it:

npx tsx <SKILL_DIR>/scripts/update-agent-settings.ts \
  --target agent \
  --agent-id "$AGENT_ID" \
  --context-window-limit 64000

Name and description

Name and description are agent-level metadata. Do not pass them with --target conversation. Values must be non-empty; the helper does not clear metadata by accident.

When the user renames you, this patch is the authoritative change — editing a name written in persona memory does not change the agent's actual name. Do both: patch the agent name here, then update any memory file that states your name so they agree.

npx tsx <SKILL_DIR>/scripts/update-agent-settings.ts \
  --target agent \
  --agent-id "$AGENT_ID" \
  --name "repo-maintainer" \
  --dry-run

npx tsx <SKILL_DIR>/scripts/update-agent-settings.ts \
  --target agent \
  --agent-id "$AGENT_ID" \
  --description "Maintains repository configuration and review-ready PRs." \
  --dry-run

Do not patch llm_config directly. Use letta model for the model and model_settings, context_window_limit for the context limit, and name/description for metadata. Then read back the agent or conversation (letta model get for model fields) and verify the result.

Compaction settings

Compaction controls how old messages are summarized when context is evicted. Bad compaction prompts cause delayed, progressive context loss as future compactions discard useful state. Good ones preserve goals, files, commands, test results, blockers, and current state.

Use the helper for prompt changes. Even --dry-run fetches current compaction settings so omitted fields are preserved in the preview. Live writes require --confirm-compaction-prompt.

npx tsx <SKILL_DIR>/scripts/update-compaction-prompt.ts \
  --prompt-file /tmp/compaction-prompt.txt \
  --mode self_compact_sliding_window \
  --clip-chars 50000 \
  --dry-run

Read references/compaction-prompt-patterns.md before drafting a new prompt.