joshuaswarren/remnic/remnic-memory-workflow

When to use

Use this skill as the default playbook whenever an agent picks up a task that could benefit from prior context, or when the user explicitly asks the agent to remember or recall something. It is the umbrella workflow — individual skills (remnic-recall, remnic-remember, remnic-search, remnic-entities, remnic-status) implement the detailed steps.

Triggers:

• The user opens a new task, ticket, or branch.
• The user says "what do you know about X", "have we talked about Y before", "remind me of Z".
• The user asks the agent to save a preference, decision, or fact.
• A long-running turn produces a durable outcome worth capturing.

Inputs

• Current user request (natural language).
• Optional: active project path, ticket number, branch name.
• Optional: topic keywords surfaced earlier in the session.

Procedure

1. Recall first. Call remnic_recall with a concise natural-language query derived from the user's request. Pull 3–8 results. Skim them for anything relevant.
2. Mention relevant memories briefly to the user if they materially change the plan, otherwise quietly use them as context.
3. Observe during work. For significant tool outputs (file edits, command exits, test results) call remnic_observe so Remnic can keep its ambient context fresh.
4. Deep search when needed. If remnic_recall misses something the user insists exists, call remnic_lcm_search with a more literal phrase.
5. Browse entities. When the user references a specific project, person, or concept by name, call remnic_entity_get to pull its facts and relationships.
6. Remember at the end. Before ending the turn, store any durable decision, preference, or finding with remnic_memory_store. Keep each entry under ~300 tokens.

Efficiency plan

• Batch recalls. One broad query is usually cheaper than several narrow ones.
• Skip recall for trivially local requests (e.g., "format this JSON").
• Reuse the current turn's recall results instead of re-querying within the same turn.
• Store each memory once. If you are about to store something you just recalled, update instead of duplicating.

Pitfalls and fixes

• Pitfall: Storing transient task state ("user is running tests now"). Fix: Only store facts with durable value — decisions, preferences, conclusions, long-lived context.
• Pitfall: Leaking secrets into memories. Fix: Redact API keys, tokens, and credentials before calling remnic_memory_store.
• Pitfall: Drowning the user in recalled context. Fix: Summarize in 1–3 bullet points; offer to expand on request.
• Pitfall: Forgetting the write. Fix: Make "remember the decision" the last step of any non-trivial turn.

Verification checklist

• [ ] Recall was attempted before the agent committed to a plan.
• [ ] Any user-facing mention of recalled context was concise and attributed.
• [ ] Durable decisions, preferences, and findings were stored via remnic_memory_store.
• [ ] No secrets, credentials, or transient state were written.
• [ ] Legacy engram_* aliases were NOT preferred over canonical remnic_* tool names.

Tool names: this skill uses the canonical remnic_* MCP tools. Legacy engram_* aliases (engram_recall, engram_memory_store, engram_lcm_search, engram_entity_get, engram_observe) remain accepted during v1.x for backward compatibility.