Capture what matters before the session ends. Save decisions, procedures, and learnings as durable memories that any connected AI tool can find later.
What to distill
- Decisions with rationale ("we chose PostgreSQL because ACID is required")
- Procedures: non-obvious steps, workarounds, setup sequences
- Learnings: surprises, gotchas, corrections to prior assumptions
- Preferences: user's stated preferences for future reference
- Plans that future sessions will need to resume
- Context that would be lost when the session ends
What to skip
- Routine code changes, obvious fixes, standard operations
- Work in progress that will change before it matters
- Generic information already widely known
- Simple Q&A answerable from documentation
Workflow
If this session already exposes the Nowledge Mem MCP server:
- Search first with
memory_searchto avoid duplicates. - If an existing memory captures the same concept, use
memory_update. - Otherwise, create it with
memory_add.
Otherwise:
- Search first to avoid duplicates:
nmem --json m search "concept" - If an existing memory captures the same concept, update it:
- Otherwise, create a new memory:
At the end of a substantial task, explicitly check whether one durable memory should be added or updated. Do not skip that review just because the user did not ask.
Unit types
fact, preference, decision, plan, procedure, learning, context, event
Additional flags
- Temporal context:
--when past(orpresent,future,timeless) - Event date:
--event-start 2026-03(for time-anchored knowledge)
Importance scale
0.8-0.9: Major decisions, important procedures, breakthroughs0.5-0.7: Useful patterns, conventions, secondary decisions0.3-0.4: Minor notes, preferences, contextual observations1.0: rare, only for critical corrections or pivotal choices
After saving
Report what was stored, which unit type was used, and why each memory was worth keeping. One strong memory is better than three weak ones.
