Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.hellofriday.ai/llms.txt

Use this file to discover all available pages before exploring further.

Memory lets a space remember what happened across sessions. Friday automatically injects the 20 most recent entries from each narrative store into agent context at the start of every session — agents don’t need to explicitly fetch memory for it to be available.

​
How memory works

Each space declares one or more memory stores. At session start, the platform injects recent entries into the agent’s system prompt as labeled blocks: 
<memory space="my-space" store="notes">
- PR review for myorg/myrepo #42 completed — 2 critical findings (2026-04-29)
- User prefers compact summaries
</memory>
Agents can also write to memory explicitly during a job, recording summaries and facts for future sessions.

​
Declaring memory stores

workspace.yml
memory:
  own:
    - name: notes
      type: short_term
      strategy: narrative
    - name: history
      type: long_term
      strategy: narrative
Memory own is a list of store objects, each with a name, type, and strategy.
FieldOptionsDescription
typeshort_term, long_termShort-term suits rolling context; long-term accumulates durable facts
strategynarrativeNarrative stores auto-inject into agent context

​
What auto-injects

The 20 most recent entries from every narrative store in the space are injected at session start. You don’t configure this per-agent — if the space has memory, every agent gets it.

​
Reading and writing memory

Agents use built-in platform tools — no MCP server required:
  • memory_save — write a new entry to a store
  • memory_read — explicitly fetch entries (for time-filtering or reading beyond the 20-entry window)
  • memory_remove — remove a stale entry by ID
In a job, an LLM agent can write a session summary to memory:
workspace.yml
agents:
  session-summarizer:
    type: llm
    description: "Summarizes the session and writes to memory."
    config:
      provider: anthropic
      model: claude-haiku-4-5
      prompt: |
        Summarize what happened in this session in 1-2 sentences.
        Write the result to memory using memory_save.
      tools:
        - memory_save

​
Keep entries terse

Memory is injected on every session — verbose entries waste context tokens and dilute signal. Rules:
  • One fact per entry, under ~100 characters
  • No preamble (“The user said that…”) — write the fact directly
  • Suffix time-sensitive entries with (YYYY-MM-DD)
Good: PR review for myorg/repo #42 — 2 critical findings (2026-04-29) Avoid: The user asked me to review a pull request and I found that there were two critical security issues in the authentication module

​
Large content: use artifact references

Don’t write large results directly to memory - they’ll bloat every future session’s context. Instead save the content as an artifact and store a short reference: Step 1 - save the artifact
The agent creates an artifact via an LLM tool call. This returns art_abc123. Step 2 - write the reference to memory
The agent calls the memory_save tool with a terse string: 
memory_save("Q1 analysis report -> art_abc123 (2026-04-29)")
Step 3 - retrieve later
In a future session, the agent reads memory (auto-injected or via memory_read), sees the artifact ID, and fetches the full content: 
artifacts_get(id="art_abc123")

​
Memory mounts

Spaces can mount memory from other spaces — useful when multiple spaces share context:
workspace.yml
memory:
  own:
    - name: notes
      type: short_term
      strategy: narrative
  mounts:
    - name: shared-kb
      source: "other-workspace-id/own/notes"
      mode: ro
      scope: space
mode is ro (read-only) or rw (read-write). Mounted stores also auto-inject into agent context.