fulcrum-recall-before-writing

Recall memory before writing

Fulcrum L2 memory stores prior decisions, outcomes, lessons. Skipping = reinventing wheel, or worse, contradicting a prior agent's decision. Always call fulcrum action exec recall_memory before novel output.

When

• About to write new module / file / function.
• About to make architectural decision (naming, schema, API shape, dependency choice).
• Writing docs describing "how things work".
• Encountered term/convention you don't recognize — memory may have defined it.
• Resuming task after break or handoff.

How

Short list of search terms drawn from task goal:

fulcrum action exec recall_memory
  workspace_id: (same as your run)
  query:        "plain english what you are about to do"
  limit:        5 or 10

Run 2-3 queries, not one:

1. Goal query: "{task goal plain english}" — e.g., "reranker tokenizer batching".
2. Path query: "{file path or dir}" — e.g., "packages/core/src/memory".
3. Concept query: "{component or pattern name}" — e.g., "FTS5 fallback", "WIP limiter".

Read every returned memory before typing. Single 200-word decision from prior run can save an hour.

Citing memories

Recalled memory changed your approach? Cite:

• Final user response: "based on prior decision M-0423, this module uses text_pair tokenization..."
• Commit message: Refs memory M-0423.
• complete_agent_run summary: include memory IDs relied on.

Creates visible reasoning chain CoS can audit.

Red flags

• Wrote new file, zero recall_memory calls → probably contradicting something.
• Memory said "don't do X" and you did X → stop, reread, justify divergence or revert.
• Zero recalls on mature project → broaden query, search by path. Zero = bad query, not empty store.

See also: start-every-task, write-memory-on-completion.