subconscious-memory

Subconscious memory discipline

Local semantic memory over MCP. The borderline cases below are where the value lives: a careless recall reuses a stale answer; a careless skip re-does work.

Recall before non-trivial work

Before research, multi-file edits, or repeating a workflow, call recall with a short task description (not the full user message).

recall(task="deploy Next.js to Vercel production", threshold=0.85)

Then branch on the outcome:

• Hit at >= 0.88: reuse answer after a quick sanity check (deps, paths, dates still apply).
• Hit below 0.88: treat as borderline. Verify the answer actually fits this task before acting on it; the phrasing matched but the intent may not.
• Miss, best similarity > 0.7: do NOT just lower the threshold. Call echo with the same task first. If the nearest entry is the same task family, recall it explicitly (or lower the threshold deliberately); if it is a different family, work fresh.
• Miss, best similarity <= 0.7: uncharted; do the work normally.

Remember after a reusable outcome

After producing something reusable (commands, config location, fix pattern):

remember(task="...", answer="...", tags=["project-name", "topic"])

Skip secrets, tokens, personal data, and answers that depend on today's file state.

Responding to a near_duplicate warning

If remember returns warning="near_duplicate", a close curated entry already exists (similarity in [0.75, 0.92]). Do not blindly add a second copy. Recall or inspect nearest_entry_id, compare it to what you were about to store, then decide:

• Merge / update: if it is the same task, forget(nearest_entry_id) and re-remember the better-phrased combined version.
• Proceed: if it is genuinely a distinct task that merely reads similar, store yours (the write already happened unless you passed skip_if_duplicate=true).

Hygiene

• Run drift_report() periodically (e.g. when reviewing cache health). For flagged entries, forget and re-remember more specific variants, or raise the recall threshold for that family.
• stats() if the user asks whether memory is helping.
• Never trust a hit above 0.85 on interpretive tasks ("extract X" vs "extract Y" read almost identically to the embedder); verify first.