Writing To Memory

Writing to Memory

Memory tool surface

Three verbs, one canonical name each across every caller (chat, FSM type: llm actions, type: atlas / type: user agents):

Same verb-first pattern for artifacts: create_artifact / get_artifact / parse_artifact / display_artifact.

Repair note: these tools were renamed from memory_save / memory_read / memory_remove (and artifacts_create / artifacts_get) in 2026-05. Installed Python agents that hard-coded the old names need a grep-and-replace — see @friday/repairing-workspaces-agents for the recipe.

Narrative is the only memory strategy

Friday memory is markdown narrative stores — append-only, auto-injected into agent prompts, queryable via list_memory_entries. There is no separate retrieval, dedup, or kv backend; if a user asks for "vector search over my notes" or "KV-style lookup", surface the limitation instead.

For everything narrative covers — preferences, standing instructions, durable facts, working notes, anything you want the agent to remember and reference next turn — declare a narrative store and you're done.

How memory is injected

At the start of every turn, the 20 most recent entries from each narrative store are injected into your system prompt as:

<memory workspace="zesty_mushroom" store="preferences">
- Always archive newsletters from substack.com (2026-04-15)
- GitHub CI failures → keep unread
</memory>

Each block is labeled with workspace and store so you know exactly what you're reading. You do not need to call list_memory_entries to access this content — it's already there.

For explicit lookup, time-filtering, or reading beyond the 20-entry window:

list_memory_entries(memoryName="preferences", since="2026-01-01T00:00:00Z", limit=50)

To remove a stale entry: delete_memory_entry(memoryName, entryId).

Store selection

Check <memory_stores> in your workspace context for what's available. Pick the store that best matches the content — do not default to notes without considering the alternatives.

type: short_term is genuinely short-term: notes you write during a session don't survive past it unless you explicitly write the same content to a long_term store. type: long_term persists across sessions until explicitly removed via delete_memory_entry. Override per-store via memory.own[].ttl: <duration> in workspace.yml when you want a specific TTL different from the type default.

If the right store doesn't exist yet, call upsert_memory_own to declare it before writing. A store must exist in workspace.yml (or the active draft) before save_memory_entry will accept it.

Entry format

Write terse entries. Memory is injected on every turn — verbose entries waste tokens and dilute signal.

Rules:

• One fact per entry. Never bundle unrelated things.
• Keep entries under ~100 chars. Longer content belongs in an artifact (see below).
• Write the fact directly — no preamble ("The user told me that...", "I noticed that...").
• Suffix time-sensitive facts with (YYYY-MM-DD).
• Never duplicate. Call list_memory_entries first if unsure whether a fact is already stored; update via delete_memory_entry + save_memory_entry rather than appending a conflicting entry.

✅ Good:

Always archive newsletters from substack.com
GitHub CI failures → keep unread
Q1 email analysis → art_abc123 (2026-04-29)
Prefers CSV over JSON for data exports

❌ Avoid:

The user mentioned that they would like newsletters from substack.com to be archived automatically in the future because they find them distracting

Large content → artifact reference pattern

Anything over ~500 chars (result sets, reports, structured data, analyses) must not go directly into memory — it will bloat every future turn's system prompt.

Step 1 — register the content as an artifact in one call.

If you're in the workspace-chat assistant (the main user-facing surface), use save_artifact for text content:

save_artifact(
  filename="q1-analysis.md",
  content="<full content>",
  title="Q1 email analysis",
  summary="..."
)
→ { success: true, id: "art_abc123", type: "file", summary: "..." }

If you're in a user agent (Python/TS SDK) calling the MCP server, use the MCP create_artifact tool, which accepts the inline data object:

create_artifact(
  data={type:"file", content:"<full content>", originalName:"q1-analysis.md"},
  title="Q1 email analysis",
  summary="..."
)
→ { id: "art_abc123", type: "file", summary: "..." }

Either way the storage layer hashes the bytes (SHA-256), sniffs the mime type from the magic bytes if you don't pass mimeType, and writes to the JetStream Object Store named by hash — identical bytes dedup automatically. No filesystem hop, no write_file → create_artifact(path) two-step.

For binary content (e.g. images returned by a tool), use the MCP create_artifact with contentEncoding: "base64" (user agent), or use run_code to write the bytes to scratch and then create_artifact({path}) (workspace-chat — save_artifact rejects binary-MIME filenames):

# MCP / user-agent path:
create_artifact(
  data={type:"file", content:"<base64>", contentEncoding:"base64", mimeType:"image/png", originalName:"chart.png"},
  ...
)

Step 2 — save a terse reference to memory:

save_memory_entry(memoryName="memory", text="Q1 email analysis → art_abc123 (2026-04-29)")

Step 3 — retrieve later: When you see → art_abc123 in an injected memory entry, call get_artifact(id="art_abc123") to load the full content on demand.

This keeps the injection window lean while making large results durable across sessions.

Memory references promote artifacts to durable

FSM-emitted artifacts default to ephemeral with a grace window (default 24h after job completion); a background sweeper deletes them once the window closes. An artifact stays durable iff something references it before the sweep: a save_memory_entry text containing the artifact ID, a display_artifact call, or an aiSummary.keyDetails[].url pointing at it.

So step 2 above isn't just bookkeeping — it's the durability decision. If you don't save_memory_entry a reference (or surface the artifact some other way), it'll be gone in 24h.

If you want something gone — intermediate working state — write the artifact and never reference it. The sweeper handles cleanup. Don't try to manage lifetimes manually.

Recent artifacts auto-inject into prompts

Per-session artifact context auto-injects into LLM prompts as <retrieved_content provenance="artifact:..." origin="..." fetched_at="..."> blocks at action start (chat: per-chat-session; FSM: per-FSM-session). The block carries each artifact's summary, not its content — the LLM sees a 1-line digest plus the artifactId. If a digest looks relevant, call parse_artifact(artifactId) to load the content.

Practical implication for summary field: when you create_artifact, write a useful 1-2 sentence summary. The LLM's later turns rely on it to decide whether to expand the artifact.

Availability

save_memory_entry, list_memory_entries, delete_memory_entry, the state_* tools, the artifacts_* tools, and webfetch are wired into every execution context with workspaceId auto-injected from the runtime scope. You never pass workspaceId — the runtime overrides it (defense in depth: a foreign workspaceId in args is replaced before the tool runs).

Authoring rule for FSM type: llm actions: do NOT redeclare these in the action's tools: array. They are auto-injected on top of any allowlist you provide. Listing save_memory_entry or create_artifact in tools: is harmless but adds noise. To genuinely lock an action down to "memory + artifacts only," declare tools: [] (empty) — the auto-injected built-ins still work, the workspace MCP catalog gets narrowed away.

Stores must be declared in workspace.yml under memory.own (or reachable via an rw mount). Undeclared stores are rejected with the list of declared ones.

User-agent footgun: ctx.tools.call(...) raises ToolCallError on validation failure (e.g. store not declared, narrative-only constraint). Never swallow with except Exception — surface the error via err() or fall through to your own retry. The autopilot family of bugs all started with a swallowed ToolCallError masking a silent write loss.