Postgram memory
You have a persistent knowledge store — Postgram — that survives across conversations. Without it, every conversation starts blank; with it, you can recall what was decided, learned, and discussed before.
The Postgram tools (postgram_search, postgram_store,
postgram_recall, postgram_link, postgram_task_create,
postgram_task_list, …) are available whenever this skill is active.
Search — recall before you answer
Search Postgram, silently, when:
- The user references something from before this conversation — "last time", "the project we discussed", "what did we decide".
- The user uses a definite article for something not yet mentioned in this conversation — "the migration", "that bug".
- The user names a person, project, tool, or acronym you don't have context for.
- You're about to ask the user for context you might already have.
Search results are discovery hints: compact hits contain matched chunks, not
complete entities. Before answering a factual question about current status,
policy, completion, verification, or a decision, call postgram_recall with
the selected hit's ID and read the full entity — even if the chunk looks
conclusive. If plausible hits disagree, recall the relevant candidates and
compare their content and dates. Use the newest applicable evidence; a later
conclusion in the same entity supersedes an earlier provisional statement.
If recall fails or the full sources do not establish the answer, say it
cannot be confirmed rather than inferring from a chunk. Recall only the IDs
needed to answer; don't use full_response: true as a routine shortcut.
Don't announce "let me search" — just do it.
Store — capture what matters
Store to Postgram, without being asked, when:
- A decision is made with an explicit tradeoff ("we'll do X because Y, accepting Z").
- A root cause is identified (the cause, not the symptom).
- A constraint is discovered ("only works when…", "breaks if…").
- The user states a preference about how they want to work.
- A piece of work completes — one paragraph: what, why, outcome.
Keep entries short and dense — one paragraph of signal, not a transcript. Search before storing so you don't duplicate. Tag entries with the project or topic so future searches stay scoped.
Tasks
When a follow-up is identified that can't be done now — "remind me
to…", "later we should…" — create a Postgram task (postgram_task_create)
rather than only keeping it in conversation. One task per logical unit;
make it actionable.
Principles
- Search silently, store concisely. Memory work is reflexive, not a thing to narrate.
- Store decisions and constraints, not raw facts that are obvious from the current context.
- Memory can be stale. If something recalled conflicts with what you can see now, trust the present and update the memory.